An infrared image light-weight identification method for end-side inspection of power transformation equipment

Abstract

The application discloses a substation equipment infrared image lightweight identification method for end-side inspection, and comprises the following steps: acquiring a substation equipment infrared image, and performing pretreatment to obtain a to-be-identified image; inputting the to-be-identified image into a lightweight backbone feature extraction network to obtain a plurality of infrared feature maps with different spatial resolutions; performing global channel statistics, local channel dependence modeling and channel weight generation on the infrared feature maps to obtain a re-calibrated infrared effective feature map; inputting the infrared effective feature map into a detection branch to perform equipment category prediction, target confidence prediction and boundary box position regression, and obtaining a candidate identification result; performing confidence screening, non-maximum suppression and coordinate reverse mapping processing on the candidate identification result, and outputting a substation equipment infrared image identification result. The method can reduce the model size and the calculation complexity while maintaining good identification precision and real-time inference ability, and is suitable for application scenarios such as unmanned aerial vehicle inspection, robot inspection and edge terminal real-time identification of a substation.

Description

Technical Field

[0001] This invention relates to the field of automatic image recognition technology for power equipment, and specifically to a lightweight infrared image recognition method for power equipment for end-side inspection. Background Technology

[0002] Substation primary equipment is a crucial component of power system operation. This equipment, including surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators, is diverse in type, widely distributed, and exhibits significant spatial structural differences. With the widespread application of intelligent substation inspection, drone inspection, inspection robots, and fixed online monitoring terminals, inspection systems need to quickly identify different types of substation equipment from a large volume of images and obtain their spatial locations within the images. This provides fundamental data for subsequent equipment management, inspection record generation, target tracking, and status analysis. Therefore, automatic image recognition technology for substation equipment has become a key component of intelligent inspection systems.

[0003] Infrared imaging can acquire thermal radiation information of equipment under low light, nighttime, or complex lighting conditions. Compared to visible light images, it has lower dependence on illumination changes, making it suitable for substation inspection scenarios. However, infrared images of substation equipment inherently suffer from weak texture details, unclear edge contours, and limited grayscale differences between the target and background. Especially under conditions of long-distance shooting or high-altitude acquisition by drones, targets such as line hardware and insulator structures occupy a small proportion in the image, easily leading to problems such as blurred target boundaries, insufficient category features, and unstable target positioning. In addition, infrared images may also contain interference such as background hot spots, metal reflections, and areas with similar grayscale levels, further increasing the difficulty of visually distinguishing between different devices.

[0004] Existing deep learning-based object detection methods have been widely used in power equipment image recognition. These methods can achieve equipment category identification and location regression through end-to-end networks, improving the efficiency of inspection image processing to some extent. However, most existing detection models are derived from general-purpose visual detection networks, resulting in complex network backbones and feature fusion structures with large parameter and computational loads, making them more suitable for platforms with sufficient computing power. For UAV onboard computing units, inspection robot control platforms, and station edge terminals, large-scale detection models are difficult to deploy stably for long periods due to limitations in power consumption, memory, heat dissipation, and real-time performance. Simply reducing the number of network layers, compressing the number of channels, or simply replacing lightweight convolutions to reduce model complexity can easily weaken the ability to represent details in infrared images, leading to a decline in the recognition performance of small-scale equipment targets.

[0005] In current infrared intelligent identification technology for substation equipment, the adaptation mechanism between infrared target feature representation and real-time deployment at the terminal side remains imperfect. Existing technologies mostly focus on improving detection accuracy by deepening or widening the network structure or introducing complex attention mechanisms, but they do not fully consider the limitations of infrared inspection terminals in terms of power consumption, storage, video memory, and inference latency, resulting in a mismatch between model recognition performance and actual deployment conditions.

[0006] Chinese patent "Intelligent Temperature Measurement Method for Substation Equipment Based on Deep Learning Algorithm" (application number: 202010926900.4) discloses a technical solution that uses inspection equipment to collect infrared thermal imaging images of substation equipment and identifies the equipment under test and its key areas through a target detection model. This solution can extract the target area of ​​the equipment from the infrared image using a deep learning model, but its technical focus is on the subsequent temperature measurement process. The target detection model used is mainly used as a pre-positioning tool and is not specifically designed for lightweight equipment identification, preservation of infrared features of small targets, and suppression of background interference.

[0007] Chinese patent "Real-time Detection and Recognition Method of Power Equipment Based on Infrared Images using Artificial Intelligence" (application number: 202011246971.6) discloses a method for acquiring infrared images of various types of power equipment using an infrared thermal imager and constructing an improved YOLOv4 model to achieve target detection and recognition of power equipment. This solution improves the detection network to address issues such as complex backgrounds, low contrast, and similar equipment appearances in infrared images, thereby enhancing the accuracy of power equipment recognition. However, this method still relies heavily on YOLOv4 and its improved networks, resulting in a relatively heavy backbone structure and detection model with a large number of parameters and computational loads. It also fails to adequately consider the deployment constraints of low-power platforms such as drones, mobile inspection terminals, and station-level edge devices.

[0008] Chinese patent "A Method for Infrared Image Recognition of Substation Equipment Based on Feature Recognition" (application number: 202211590327.X) discloses a method for recognizing multiple types of substation equipment by utilizing interest feature extraction and target detection network training to address the problem of difficulty in distinguishing equipment from the background in infrared images of substation equipment. This scheme emphasizes enhancing the representation of target regions through interest features, which has a certain effect on distinguishing equipment targets from the background. However, it still relies heavily on pre-designed or extracted interest features, and does not adequately consider the end-to-end lightweight representation capability of infrared targets at different scales. Furthermore, it does not further construct a low-overhead channel response enhancement mechanism to improve the representation capability of effective infrared equipment features. Summary of the Invention

[0009] To address the shortcomings of existing infrared image recognition methods for substation equipment in end-side inspection scenarios, such as large model parameters, high computational complexity, insufficient representation of weak infrared texture features, poor stability in identifying small-scale equipment targets, and inadequate suppression of background thermal interference, this invention proposes a lightweight infrared image recognition method for substation equipment in end-side inspection scenarios. Under single-modal input conditions of infrared images, a lightweight backbone feature extraction network is combined with a low-overhead channel response recalibration mechanism. Infrared features at different scales are extracted through multi-level lightweight inverse residual feature units, and a channel-free attention structure is used to enhance the effective infrared target channels while suppressing background hotspots, weak texture regions, and ineffective response channels. This reduces the number of model parameters and computational load while improving the real-time performance and stability of infrared image recognition for substation equipment.

[0010] The technical solution adopted in this invention is as follows: A lightweight infrared image recognition method for substation equipment for end-side inspection includes the following steps: Step 1: Acquire infrared images of the power equipment, preprocess the infrared images to obtain images to be identified that meet the input requirements of the lightweight backbone feature extraction network; Step 2: Construct a lightweight backbone feature extraction network. Input the image to be identified obtained in Step 1 into the lightweight backbone feature extraction network to obtain multiple infrared feature maps with different spatial resolutions. Step 3: Construct a channel response recalibration module at the output of the lightweight backbone feature extraction network to perform global channel statistics, local channel dependency modeling, and channel weight generation on the infrared feature map obtained in Step 2, and obtain the recalibrated effective infrared feature map. Step 4: Input the effective infrared feature map obtained in Step 3 into the detection branch to perform device category prediction, target confidence prediction and bounding box position regression to obtain candidate recognition results; Step 5: Perform confidence screening, non-maximum suppression, and coordinate inverse mapping on the candidate recognition results obtained in Step 4, and output the infrared image recognition results of the substation equipment.

[0011] It also includes step 6: Constructing a lightweight infrared image recognition system for power equipment, deploying the recognition processes corresponding to steps 1 to 5 on the end-side inspection platform to realize infrared image acquisition, preprocessing, lightweight recognition, result output and inspection data management.

[0012] Step 1, which involves acquiring and preprocessing infrared images of the power equipment, specifically includes: S1.1: Collect infrared images of substation equipment during inspection. Each infrared image must contain at least one type of substation equipment target, including surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators. Include the substation equipment's category number, Chinese category, English category, and typical target characteristics. S1.2: The acquired raw infrared images are formatted and read to obtain the raw infrared image matrix; if the raw infrared image is a temperature matrix, the temperature values ​​are mapped to a grayscale image; if the raw infrared image is already a pseudo-color or grayscale image, its grayscale intensity information or thermal radiation intensity information is extracted as subsequent recognition input. Specifically: Let the original infrared temperature matrix obtained be: (1); In equation (1), Represents spatial coordinates in the original infrared image The corresponding temperature value at that location; and These represent the pixel coordinates in the width and height directions of the image, respectively; and These represent the height and width of the original infrared image, respectively; express OK The space of real matrices in the column; When the input data is a temperature matrix, the temperature values ​​are linearly mapped to a preset grayscale range based on the temperature value range. The calculation method is as follows: (2); In equation (2), Represents the pixel coordinates after temperature value mapping The grayscale value at that location; This indicates the number of gray levels. When using an 8-bit grayscale image... ; and These represent the lower and upper temperature limits used when performing grayscale mapping, respectively, which are determined based on the lowest and highest temperatures of the current infrared image, or preset based on the effective temperature measurement range of the infrared imaging device. This indicates an extremely small positive number that prevents the denominator from being zero; This represents a cutoff function used to limit temperature values ​​that exceed a preset temperature range. Within the range.

[0013] When the input infrared image is a three-channel pseudo-color image, let the pixel values ​​of its red, green, and blue channels be respectively... , and Then it can be converted into a single-channel grayscale intensity: (3); In equation (3), , , Let these represent the grayscale conversion weights corresponding to the red, green, and blue channels, respectively, and satisfy the following conditions: (4); If the original input is already a single-channel grayscale infrared image, then directly set: (5); In equation (5), Represents the original grayscale infrared image in pixel coordinates The pixel intensity value at that location.

[0014] Step 1.3: Perform size normalization on the infrared images, scaling input images of different resolutions to a preset network input size. During scaling, record the original infrared image size, scaling ratio, and fill pixel range so that the bounding box coordinates output by subsequent detection branches can be mapped back to the original infrared image. Specifically: Let the original infrared image size be: (6); The default network input size is: (7); in, Enter the height for the network. Enter the network width; To avoid distortion of the device's geometric proportions caused by direct stretching, the image is scaled proportionally to its original aspect ratio. Represented as: (8); This is the scaling factor for the image; This indicates a minimum value operation, ensuring that the scaled image does not exceed the network input size and avoids geometric distortion.

[0015] The scaled image dimensions are represented as follows: (9); (10); In the formula, and These represent the width and height of the image after proportional scaling, respectively. This represents the rounding operation that takes the nearest integer.

[0016] The total number of pixels that need to be filled in the width and height directions of the network input image are expressed as follows: (11); (12); In the formula, This indicates the total number of pixels that need to be filled in the width direction of the network input image; This represents the total number of pixels that need to be filled in the height direction of the network input image.

[0017] When using a centrally symmetrical fill method, the fill amounts on the left and top are denoted as follows: and , is represented as: (13); In the actual pixel filling process, when or When the number is odd, the remaining pixels are distributed to both sides of the image by rounding down and rounding up respectively.

[0018] For any pixel coordinate in the original image After being scaled and filled proportionally, its coordinates in the network input image are... Represented as: (14); (15); In the formula, Represents the pixel coordinates in the original infrared image; This represents the coordinates of the network input image after size normalization; Indicates the image scaling factor; and These represent the fill offsets in the width and height directions, respectively.

[0019] The parameter set is recorded in step 1.3: (16); This represents the set of mapping parameters, recording the original size, network input size, scaling ratio, and padding offset, which are used in step 5 to reverse map the predicted bounding box under the network input size to the original infrared image coordinates.

[0020] S1.4: Perform grayscale normalization processing on the infrared image to normalize the image pixel values ​​or temperature mapping values ​​to a uniform range, as follows: Let the infrared image after temperature grayscale mapping or grayscale reading be... Its grayscale value range is Then the normalized pixel value Represented as: (17); In equation (17), , These represent the minimum and maximum gray values ​​of the image before normalization, respectively. Represents the normalized pixel coordinates The pixel value at that location has the following range: (18); and These represent the minimum and maximum gray values ​​of the image before normalization, respectively. This represents the smallest positive number that prevents the denominator from being zero.

[0021] When the input image is represented using 8-bit grayscale, and the grayscale value range is fixed. When the above formula simplifies to: (19); S1.5: Perform noise suppression and edge preservation processing on the infrared image: To address random noise, localized hotspots, and weakly textured areas in infrared images, median filtering, bilateral filtering, or lightweight smoothing techniques are employed to reduce noise interference while preserving the edge contours of the substation equipment and the target structural information. Specifically, this includes: Let the input infrared image be The infrared image after noise suppression processing is ,in: This represents the pixel coordinates in an infrared image.

[0022] ①: When using median filtering, the filtered pixel values Represented as: (20); In equation (20): Represented by pixels A local neighborhood window centered on the target; This represents the pixel coordinates within the local neighborhood window; This indicates that the median grayscale value of the pixels within the local neighborhood window is taken; This represents the pixel value after median filtering; Represents neighboring pixels The grayscale value at that location.

[0023] ②: When using bilateral filtering, the filtered pixel values As shown in formula (21): (twenty one); In equation (21): This represents the pixel value after bilateral filtering; Represents the normalization coefficient; This represents the spatial distance weighting function. This represents the gray-level similarity weighting function; Represents the two-dimensional Euclidean distance. Represents the center pixel coordinates The grayscale value or pixel intensity value at that location.

[0024] Among them, the normalization coefficient As shown in formula (22): (twenty two); The spatial distance weight function and the gray-level similarity weight function are respectively expressed as: (twenty three); (twenty four); In the formula, Represented by natural constant An exponential function with base 0; This represents the spatial domain standard deviation, used to control the impact of the spatial distance between neighboring pixels on the filter weights; This represents the standard deviation of the grayscale domain, used to control the impact of grayscale differences between neighboring pixels on the filtering weights. This represents the squared spatial distance between the center pixel and its neighboring pixels. This represents the squared grayscale difference between the center pixel and its neighboring pixels.

[0025] ③: When using the lightweight smoothing method, the filtered pixel values As shown in equation (25): (25); In equation (25): Represents a local neighborhood window The number of pixels within. This represents the pixel value after lightweight smoothing (mean filtering). S1.6: During the training phase, the device category labels and bounding box coordinates in the annotation file are synchronously transformed to correspond to the preprocessed network input image; specifically as follows: During the inference phase, only the image scaling and padding parameters are retained for coordinate inverse mapping in step 5.

[0026] Let the first part of the original infrared image be... The category label for each target is Its actual bounding box is represented as: (26); In equation (26), and They represent the first The pixel coordinates of the top left and bottom right corners of the target bounding box in the original infrared image. Indicates the first The true bounding box of each target; Based on the scaling ratio obtained in step 1.3 and fill offset , The coordinates of the bounding box after synchronous transformation are: (27); (28); (29); (30); In the above formula, , These represent the coordinates of the top left and bottom right corners of the bounding box in the network input image after synchronous transformation; Accordingly, the bounding box after synchronous transformation is represented as: (31); This represents the bounding box after synchronous transformation; The equipment category label remains unchanged during geometric transformation: (32); This represents the category label after the geometric transformation, and the category label remains unchanged before and after the transformation. When the training annotations use normalized center point coordinates and width / height, further calculations can be performed: (33); (34); (35); (36); In the formula, They represent the first The normalized coordinates of the center point of each target bounding box relative to the width and height of the network input image; and These represent the normalized bounding box width and height, respectively; and These represent the width and height of the input image to the network, respectively.

[0027] Step 2, which involves constructing a lightweight backbone feature extraction network, specifically includes: S2.1: Based on the single-stage target detection framework, a lightweight backbone feature extraction network for infrared images is constructed. The lightweight backbone feature extraction network consists of an initial convolutional layer and multiple levels of lightweight inverse residual feature units, which are used to extract multi-scale features from the input infrared image, from shallow edge texture to deep semantic information.

[0028] S2.2: Input the image to be identified obtained in step 1 into the initial convolutional layer to obtain the initial infrared feature map. The initial convolutional layer is used to complete the mapping of the infrared image from pixel space to feature space, and to initially extract information on the thermal edges, local textures, and regional grayscale differences of the power equipment. Specifically: Let the network input infrared image obtained in step 1 be: (37); In the formula, This represents the infrared image tensor input to the network. Indicates the number of input channels; , These represent the height and width of the network input image, respectively. The initial infrared feature map is obtained through the initial convolutional layer. , is represented as: (38); In the formula, This represents the infrared feature map output by the initial convolutional layer; Represents a non-linear activation function; This indicates a batch normalization operation; Indicates the initial convolutional layer kernel parameters; * represents the convolution operator. This represents the bias parameter.

[0029] Furthermore, the first The output channels are located in space. The eigenvalues ​​at that location are represented as: (39); In the formula, Represents the initial feature map. The output channels are located in space. Eigenvalues ​​at; Indicates the output channel number; Indicates the input channel number; Indicates the initial convolution kernel size; , Indicates the position index (row or column direction) inside the convolution kernel; Indicates the connection of the first Input channel and the first Output channel, convolution kernel The weighting parameters of the location; The first input tensor represents the... One channel; Indicates the convolution stride; Indicates the fill size; Indicates the first The bias corresponding to each output channel.

[0030] S2.3: The initial infrared feature map is input into a multi-level lightweight inverse residual feature unit to sequentially obtain infrared feature maps with different spatial resolutions. The shallow high-resolution feature map is used to preserve device edges and small target textures, the mid-level semantic feature map is used to express the device's structural morphology, and the deep low-resolution feature map is used to provide semantic information about the device category. Details are as follows: Let the initial convolutional layer output be , No. The lightweight feature extraction unit is denoted as The multi-level feature extraction process can then be represented as: (40); In equation (40), Indicates the number of levels in a lightweight feature extraction network; and They represent the first The input and output feature maps of the level feature extraction unit; Indicates the first Lightweight feature transformation.

[0031] This represents the output of the initial convolutional layer, i.e., the 0th level feature; Indicates a network level index; Three scale features were selected from different network depths and denoted as follows: (41); (42); (43); In the formula, , and These represent infrared feature maps at three scales with step sizes of 8, 16, and 32, respectively. , and These represent the number of channels in the corresponding scale feature map; and This indicates the height and width of the input image to the network.

[0032] When the network input size is At that time, the corresponding spatial dimensions are as follows: (44); Its multi-scale output modes of P3 / 8, P4 / 16, and P5 / 32 correspond to the three detection scale settings of the YOLOv11 detection structure.

[0033] S2.4: The lightweight inverse residual feature unit comprises a channel expansion layer, a spatial modeling layer, a channel projection layer, and a residual connection layer. The channel expansion layer uses 1×1 convolutions to expand the input features, enhancing the feature representation dimension. The spatial modeling layer uses depthwise convolutions or depthwise separable convolutions to extract local infrared spatial features. The channel projection layer uses 1×1 convolutions to compress the number of channels, reducing computational cost. When the input and output feature sizes and number of channels are consistent, the residual connection layer adds the input features to the projected features to enhance gradient propagation and reduce information loss caused by lightweight compression. Details are as follows: Let the input features of the lightweight reciprocal residual feature unit be: (45); In the formula, This represents the input features of the lightweight inverse residual feature unit; This represents the number of channels for the input feature; , These represent the height and width of the feature map, respectively.

[0034] First, optional channel-wise spatial feature extraction is performed: (46); In the formula, This indicates channel-wise convolution (depth convolution), where each channel is spatially convolved independently. , These represent the kernel size and stride of the preceding channel-wise convolution, respectively.

[0035] Then adopt Channel expansion via pointwise convolution: (47); In the formula, This indicates the characteristics after channel expansion; Indicates the use of channel expansion Pointwise convolution; The expanded number of channels is expressed as: (48); In equation (48), Indicates the channel spread factor; This indicates the number of feature channels after expansion; This indicates the rounding operation.

[0036] Then perform channel-wise spatial convolution: (49); In the formula, This represents the intermediate features obtained by channel-wise spatial convolution after channel expansion; , These represent the kernel size and stride of the channel-wise spatial convolution after channel expansion, respectively. Then through Pointwise convolution completes channel projection: (50); In the formula, Indicates Features obtained after pointwise convolution channel projection (compression); This represents pointwise convolution used for channel projection; When the input and output feature spaces have the same size and the same number of channels, residual connection is used: (51); In the formula, This represents the final output feature of the lightweight inverse residual feature unit; otherwise: (52); When the input / output dimensions or number of channels are inconsistent, residual connections are not enabled, and the projected features are output directly.

[0037] S2.5: The lightweight backbone feature extraction network adopts a MobileNetV4-type backbone structure and selects a convolutional small model configuration as the backbone parameter configuration. By decomposing standard convolution operations through channel-wise spatial convolution and point-wise channel-wise convolution, infrared spatial texture extraction and cross-channel information fusion are performed separately. Specifically: In standard convolution, the first The eigenvalues ​​of each output channel are represented as follows: (53); In equation (53), Represents the first... One channel; The first feature map is the one whose output feature map is represented by the first feature map. One channel; Indicates the standard convolution kernel parameters; Indicates the number of input channels; This represents the side length of the convolution kernel.

[0038] When using depthwise separable convolution, spatial convolution is first performed on each input channel separately: (54); In the formula, Indicates the first Intermediate features obtained by channel-wise spatial convolution of each input channel; Indicates the first The depthwise convolution kernel parameters corresponding to each channel; Then adopt Pointwise convolution completes cross-channel information fusion: (55); In the formula, Indicates by the first The intermediate channel is mapped to the first... Each output channel Pointwise convolution parameters.

[0039] Therefore, the standard convolution parameter count is: (56); In the formula, The number of parameters representing a standard convolution; Indicates the number of channels in the output feature map; This represents the number of spatial parameters of a single convolution kernel.

[0040] The number of parameters in a depthwise separable convolution is: (57); In the formula, The number of parameters representing a depthwise separable convolution; the first term The second term represents the number of parameters in the channel-wise spatial convolution; express Pointwise convolution parameters.

[0041] S2.6: Let the number of input feature map channels be... The number of channels in the output feature map is The kernel size is The feature map space size is Standard convolution computational cost Depthwise separable convolution The computational load is expressed as follows: (58); (59); In the formula, and These represent the height and width of the output feature map, respectively; and These represent the number of channels in the input feature map and the output feature map, respectively. Indicates the side length of the convolution kernel; Indicates the computational cost of standard convolution; The computational cost of depthwise separable convolution is expressed as follows. Equation (58) represents the computational cost of standard convolution, and Equation (59) represents the computational cost of depthwise separable convolution. In Equation (59), the first term represents the computational cost of channel-wise spatial convolution, and the second term represents... The computational cost of cross-channel information fusion using pointwise convolution.

[0042] S2.7: To ensure that the output features of the backbone network can be adapted to subsequent detection branches, the infrared feature maps of different scales output by the multi-level lightweight inverse residual feature units are numbered according to scale, and are designated as P3, P4 and P5. Among them, P3 retains higher resolution and more detailed information, and is used for small-scale substation target recognition; P4 takes into account both spatial information and semantic expression; P5 contains stronger semantic information and is used for larger-scale equipment targets and overall structure recognition.

[0043] Step 3 specifically includes: S3.1: Input the infrared feature map of at least one scale obtained in step 2 into the channel response recalibration module. The channel response recalibration module is set between the output end of the lightweight backbone feature extraction network and the input end of the detection branch. It adaptively evaluates the importance of different channels in the infrared feature map and enhances the target-related feature response of the device without significantly increasing the number of parameters and computation.

[0044] S3.2: For any scale of input infrared feature map ,in: Indicates the number of channels. and These represent the height and width of the feature map, respectively. The channel response recalibration module processes the data using a parallel approach that preserves the original feature bypass and generates channel weight branches. Input feature map On the one hand, it is directly passed along the bypass branch to preserve the original spatial structure information; on the other hand, it is input into the global average pooling layer (GAP) to extract the global response statistics of each channel. For the Each channel undergoes global average pooling, and its channel descriptor value Represented as: (60); In equation (60), Indicates the input infrared feature map The Middle The spatial location of each channel eigenvalues ​​at This represents the spatial position index along the height direction of the feature map, with a value range of [value range missing]. ; This represents the spatial position index along the width direction of the feature map, with a value range of [value range missing]. ; Indicates the first The channel description values ​​obtained after global average pooling of each channel; Represents the total number of spatial locations in the feature map S3.3: Obtaining the channel description vector Then, based on the number of channels in the input feature map Adaptive determination of one-dimensional convolution kernel size , The convolution kernel size With the number of channels Correspondingly, infrared feature maps of different channel sizes employ matched local channel interaction ranges. The adaptive convolutional kernel size is expressed as: (61); In equation (61), and Preset adjustment parameters; This indicates taking the nearest odd number; Indicates the size of a one-dimensional convolutional kernel (local channel interaction range); Indicated by the number of channels The adaptive convolution kernel size mapping function is the independent variable; This indicates the number of channels in the input feature map; Represents the logarithmic function with base 2; S3.4: The one-dimensional convolution kernel size determined using S3.3 , channel description vector Perform one-dimensional convolution operations to establish local dependencies between adjacent channels and obtain channel interaction response vectors. The calculation process is expressed as follows: (62); In equation (62), Indicates the kernel size as One-dimensional convolution operation.

[0045] Represents the channel interaction response vector; This represents the channel description vector composed of all channel description values. ; S3.5: The channel interaction response vector obtained in S3.4 Inputting the Sigmoid activation function yields the channel weight vector. The calculation process is expressed as follows: (63); In equation (63), This represents the Sigmoid activation function; This represents the generated channel weight vector.

[0046] S3.6: The channel weight vector obtained in S3.5 Compared with the original infrared feature map preserved by the bypass in S3.2 Perform channel-by-channel multiplication to obtain the recalibrated effective infrared feature map. The calculation process is expressed as equation (64): (64); In the formula, Indicates the first The channel weights corresponding to each channel; Indicates the first Output characteristics after recalibration of each channel; The original feature map representing bypass preservation. One channel.

[0047] Or expressed as equation (65): (65); In the formula, Indicates the first Output characteristics after recalibration of each channel; This indicates a channel-by-channel multiplication operation.

[0048] S3.7: For infrared feature maps of different scales, execute S3.2~S3.6 respectively to independently generate channel weights for the corresponding scales and complete feature recalibration. When the input feature map is a shallow high-resolution feature, the channel response recalibration module focuses on enhancing device edges, small target textures, and local thermal structure information; when the input feature map is a mid-level feature, it focuses on enhancing device morphology and local structural information; when the input feature map is a deep low-resolution feature, it focuses on enhancing semantically relevant channels for device categories.

[0049] S3.8: Output the recalibrated infrared effective feature maps at each scale to the subsequent detection branch.

[0050] Step 4 specifically includes: S4.1: Input the multi-scale infrared effective feature maps obtained in S3 into the single-stage detection branch, which specifically includes a classification prediction sub-branch, a target confidence prediction sub-branch, and a bounding box regression sub-branch. The classification prediction sub-branch is used to determine the equipment category of the candidate target, the target confidence prediction sub-branch is used to determine whether there is a substation equipment target in the candidate region, and the bounding box regression sub-branch is used to predict the location of the equipment target in the image.

[0051] The classification prediction sub-branch outputs the probability value of each candidate target belonging to the categories of surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators. The target confidence prediction sub-branch outputs the probability that the candidate region contains substation equipment targets. The bounding box regression sub-branch outputs the coordinates of the bounding box center point, width, and height. Details are as follows: For the Each scale feature map The output of the detection branch is represented as: (66); In equation (66), This indicates three detection scales; Indicates the first Each scale corresponds to a detection head; This represents the bounding box regression output; This represents the target confidence prediction output; This indicates the category prediction output.

[0052] For feature map location First The probability of a device class is expressed as follows: (67); in, (68); In the formula, Indicates the output of the detection head. Class prediction value; Indicates the probability of the corresponding category; ,in This indicates the total number of equipment categories.

[0053] The predicted device category is: (69); In the formula, Indicates the location of the feature map Predicted device category at the location; Represents the probability of taking the class. Largest category number .

[0054] The corresponding target confidence level is defined as: (70); In the formula, Indicates the location of the feature map The predicted confidence level of the target at the location; This represents the Sigmoid activation function; This indicates the confidence level that the corresponding candidate region contains the target of the power equipment.

[0055] Suppose that the bounding box regression yields the predicted bounding box: (71); This can then be further converted into a center point and width / height format: (72); (73); (74); (75); In the formula, and These represent the coordinates of the top-left and bottom-right corners of the predicted bounding box, respectively. Indicates the coordinates of the center point of the bounding box; and These represent the width and height of the bounding box, respectively.

[0056] Therefore, the first The candidate identification results are uniformly represented as follows: (76); In equation (76), , Indicates the first One candidate identification result; Indicates the predicted device category; Indicates the confidence level of the target; Indicates the coordinates of the predicted bounding box center point; and These represent the width and height of the predicted bounding box, respectively.

[0057] For each predicted target, the detection branch outputs its class probability, target confidence, and bounding box coordinates. Each scale feature map corresponds to three types of outputs: classification prediction, confidence prediction, and bounding box regression.

[0058] S4.2: During the training phase, the lightweight backbone feature extraction network, channel response recalibration module, and detection branch are jointly trained based on the category labels and bounding box labels in the infrared image dataset of substation equipment.

[0059] Training losses include classification loss, object confidence loss, and bounding box regression loss: The detection branch employs a joint loss function to optimize classification prediction, target confidence prediction, and bounding box regression. Specifically, the classification loss constrains the consistency between the predicted and ground truth classes, the target confidence loss determines whether the candidate region contains the substation target, and the bounding box regression loss constrains the degree of overlap between the predicted and ground truth bounding boxes. Let the predicted bounding box be The actual labeled bounding box is The intersection and union ratio of the two Represented as: (77); In equation (77), This represents the intersection area of ​​the predicted bounding box and the ground truth bounding box. This represents the area of ​​the union of the two. Based on the intersection-union ratio, the bounding box regression loss... As shown in equation (78): (78); Total loss function for detection branches As shown in equation (79): (79); In equation (79), Represents classification loss, Indicates the target confidence loss. This represents the bounding box regression loss; , , These represent the weighting coefficients of each loss term. Using the joint loss function described above, the detection branch can improve the matching degree between the predicted bounding box and the real target region while simultaneously identifying the device category. S4.3: During the inference phase, the detection branch generates candidate recognition boxes based on the infrared effective feature map output in step 3, and obtains the device category, confidence level and bounding box coordinates of each candidate target.

[0060] Step 5 specifically includes: S5.1: Read the candidate recognition results obtained in 4.3. The candidate recognition results include the device category, target confidence level, and bounding box coordinates of the candidate target. Establish a candidate box set for all candidate recognition results in the same infrared image, denoted as: (80); In equation (80), Represents the set of candidate recognition results. Indicates the first One candidate identification result, This indicates the number of candidate identification results.

[0061] S5.2: Confidence screening of candidate identification results: Let the first The target confidence level for each candidate identification result is: The preset reliability threshold is ,when When, delete the candidate recognition result; when If the candidate identification result is not found, it will be retained for subsequent processing.

[0062] S5.3: The candidate identification results after confidence level screening are grouped according to equipment category. For different categories such as surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators, candidate box sorting and duplicate box suppression are performed separately, as follows: Let the set of candidate identification results after confidence level filtering be: (81); In the formula, This represents the set of candidate identification results after confidence level filtering; This indicates the total number of candidate boxes retained after filtering; Among them, the The candidate identification results are represented as follows: (82); In the formula, Indicates the first Predicted bounding boxes of candidate targets; Indicates the predicted device category; Indicates the confidence level of the target; This represents the total number of candidate boxes retained after confidence level filtering.

[0063] For the For each device class, create a corresponding set of candidate frames: (83); In the formula, This indicates the total number of categories of power equipment. Indicates the first The set of candidate boxes corresponding to the device class; For sets Candidate results in the data are ranked according to confidence level. Sort in descending order. Let the sorted index sequence be: (84); In the formula, Indicates the first A sorted index sequence of candidate boxes arranged in descending order of confidence. This indicates an operation that returns the index of the sorted element; The negative sign indicates that the order is descending (from highest to lowest confidence level). The sorted candidate box set is then represented as: (85); In the formula, This represents the sorted set of candidate boxes; Represents the original index of the 1st, 2nd...th position in the sorted sequence; Indicates the first The number of candidate boxes corresponding to the device class.

[0064] And satisfy: (86); In the formula, Indicates the first The number of candidate boxes corresponding to the device class; This indicates an operation that returns the index of the sorted element.

[0065] S5.4: Perform non-maximum suppression processing on candidate bounding boxes within the same device category. Specifically: First, the candidate boxes are sorted from highest to lowest target confidence level, and the candidate box with the highest confidence level is selected as the current retained box; the details are as follows: Let the first The current set of candidate boxes to be processed is ,in The first nonmaximum suppression represents the... In the next iteration, the index of the candidate box with the highest confidence is represented as: (87); The corresponding candidate results Add to the final reserved set : (88); In the formula, Indicates the first The index of the candidate box with the highest target confidence in the next iteration; Indicates the first The target confidence level of each candidate box; Indicates the first The first iteration has already retained the... A collection of candidate boxes; This represents the set union operation (combining the current best box into sets). (Add to the reserved set) This represents the updated set to be retained.

[0066] Then, the intersection-union ratio (IUR) between the remaining candidate boxes and the currently retained box is calculated. If the IUR is greater than a preset threshold, it is considered that the two correspond to the same substation equipment target, and the candidate boxes with lower confidence are judged as duplicate candidate boxes and deleted; specifically as follows: Let the currently retained box be: (89); In the formula, This indicates the currently reserved box (the box with the highest confidence level); , This indicates the coordinates of its top-left and bottom-right corners.

[0067] The candidate boxes to be compared are: (90); In the formula, Indicates the candidate boxes to be compared; , Indicates the coordinates of its top left and bottom right corners; The width and height of the intersection region are represented as follows: (91); In the formula, This indicates the width of the intersection area between the two frames; : The right boundary of the intersection (take the smaller of the right boundaries of the two boxes); Indicates the left boundary of the intersection (the larger of the left boundaries of the two boxes); outer layer This indicates that the intersection width is 0 when the two boxes do not intersect.

[0068] (92); In the formula, Indicates the height of the intersection area of ​​the two frames; The area of ​​intersection is: (93); In the formula, This represents the area of ​​the intersection of the two frames.

[0069] The areas of the two candidate boxes are as follows: (94); In the formula, This indicates the area of ​​the currently reserved box.

[0070] (95); In the formula, This represents the area of ​​the candidate boxes to be compared.

[0071] The intersection, union, and ratio of the two can be expressed as: (96); In the formula, This indicates the intersection-union ratio (IoU) of the reserved box and the box to be compared. Let the nonmaximum suppression threshold be... ,when: (97); and: (98); In the formula, Indicates the confidence level of the current reserved box; Indicates the confidence level of the candidate boxes to be compared; Then the candidate box Determined as the currently retained frame The corresponding duplicate prediction boxes are then removed from the set to be processed.

[0072] The candidate box set after one iteration is represented as: (99); In the formula, This represents the set of candidate boxes remaining after one iteration. To perform a set difference operation, first remove the currently retained box. ;condition This means that only candidate boxes whose intersection-union ratio with the currently retained boxes does not exceed the threshold are retained, and duplicate boxes are deleted.

[0073] Repeat the above process until: (100); In the formula, This indicates an empty set, meaning the iteration terminates when the set to be processed is empty; at this time, That is, the first The set of candidate recognition results retained by the device after nonmaximum suppression.

[0074] Repeat the above process until all candidate boxes in this category have been processed. S5.5: Perform coordinate inverse mapping on the bounding boxes retained after nonmaximum suppression: Based on the image scaling ratio, fill size, and original image size recorded in step 1, the bounding box coordinates under the network input size are converted back to the original infrared image coordinates to obtain the actual position of the device target in the original image. Specifically: Let the coordinates of the predicted bounding box retained after nonmaximum suppression in the network input image be: (101); In the formula, This represents the coordinates of the predicted bounding box preserved by NMS in the coordinate system of the network input image; , This represents the coordinates of the top-left and bottom-right corners of the predicted bounding box in the network input image.

[0075] Based on the image scaling ratio recorded in step 1.3 And the fill offset in the horizontal and vertical directions. and The predicted bounding box is then mapped inversely to the original infrared image coordinate system, as follows: (102); In the formula, This represents the predicted bounding box in the coordinate system of the network input image; This represents the bounding box after inversion onto the original infrared image; and These represent the coordinates of the top-left and bottom-right corners of the predicted bounding box in the network input image, respectively. This represents the scaling factor of the image recorded in step 1.3; and These represent the fill offsets of the image in the horizontal and vertical directions, respectively.

[0076] S5.6: Output the final infrared image recognition results of the substation equipment: The recognition results include at least the device category, bounding box coordinates, target confidence score, and image number; Depending on the actual inspection needs, it also includes the collection time, inspection point, equipment number, collection terminal number, and identification result storage path.

[0077] S5.7: Associate the final recognition result with the original infrared image. For scenarios requiring visualization, overlay the device category name, target confidence score, and bounding box position onto the original infrared image to form a visual recognition map; specifically as follows: For scenarios requiring data retrieval, the recognition results can be output to the inspection management platform in the form of structured data.

[0078] Let the first The final identification result is: (103); In the formula, Indicates the first The final identification result is a triple: bounding box, category, and confidence score.

[0079] in: (104); Indicates the final bounding box; Indicates the equipment category; Indicates the confidence level of the target.

[0080] Definition of the first The set of boundary pixels of each bounding box is: (105); In the formula, Indicates the first The set of bounding box pixels (i.e., all pixels on the four sides of the rectangle). The first set: the top and bottom horizontal edges ( Pick or , (takes values ​​continuously between the left and right boundaries). The second set consists of two perpendicular edges, left and right. It represents the union of sets.

[0081] For all final bounding boxes, define a visual boundary mask: (106); In the formula, Represents a visual boundary mask (binary mask); This represents the union of all final bounding box pixels; The image after the bounding boxes are superimposed is represented as: (107); In equation (107), Represents the original infrared image; This represents a visualized image after the recognition results have been overlaid. Represents the bounding box mask; Indicates the relationship with the first The preset display value or display color corresponding to this type of device.

[0082] No. The visual text labels corresponding to each target are represented as follows: (108); In equation (108), This represents the visual text label corresponding to the i-th target; Indicates based on category number The device category name obtained from the query; This indicates the target confidence level after formatting according to a preset number of decimal places; This indicates a string concatenation operation.

[0083] Text Label It can be set near the top left corner of the bounding box: (109); In the formula, Represents text label The display position coordinates are set at the top left corner of the bounding box. nearby.

[0084] This results in a visual recognition outcome that includes the device category name, target confidence level, and spatial location.

[0085] S5.8: The data fields of the final recognition result are set according to a table. The table includes field names, field meanings, and data types, which are used to unify the storage format and interface output format of the recognition results.

[0086] In step 6, a lightweight infrared image recognition system for substation equipment is constructed and deployed on the end-side inspection platform, specifically including: S6.1: An infrared image input interface is set in the end-side inspection platform. The infrared image input interface is connected to the UAV infrared camera, the inspection robot infrared imaging equipment, the fixed infrared monitoring terminal or the handheld infrared imaging equipment to receive infrared images collected during the substation inspection process and send the infrared images into the subsequent image processing flow.

[0087] S6.2: Establish a preprocessing procedure for the input infrared image. The preprocessing procedure includes image size normalization, temperature grayscale mapping, noise suppression, pixel normalization, and coordinate mapping parameter recording. After preprocessing, the original infrared image is converted into a recognition image that meets the input requirements of the lightweight recognition network, while retaining the image scaling ratio and padding information so that the subsequent recognition results can be mapped back to the original image coordinates.

[0088] S6.3: The image to be identified is fed into the lightweight feature extraction unit, which performs the backbone feature extraction process in step 2 to obtain infrared feature maps with different spatial resolutions. This unit reduces the number of parameters and computational load during edge-side inference through a lightweight backbone network, providing a multi-scale feature foundation for subsequent device target recognition.

[0089] S6.4: After the lightweight feature extraction unit, a channel response recalibration unit is set up to perform the channel weight calculation and feature recalibration processing in step 3 on the multi-scale infrared feature map. Through this processing, the channels related to the device's thermal edge, main body contour, and structural texture are enhanced, while the channel responses corresponding to background hot spots and weakly correlated areas are suppressed, thereby obtaining an effective infrared feature map.

[0090] S6.5: The effective infrared feature map is input into the recognition output unit, which performs candidate box generation, confidence level filtering, non-maximum suppression, and coordinate inverse mapping processing, and outputs the infrared image recognition result of the substation equipment. The recognition result includes equipment category, equipment location, and target confidence level.

[0091] S6.6: Store the equipment category, bounding box coordinates, target confidence score, inspection task number, acquisition time, inspection point, equipment number, and original infrared image. For scenarios requiring manual verification, the recognition box, category name, and confidence score can be overlaid on the original infrared image to generate a visual recognition result; for scenarios requiring platform access, the recognition result can be converted into structured data and output to the inspection management platform.

[0092] S6.7: During the operation of the end-side inspection platform, the infrared image input interface receives inspection images in real time. The preprocessing process generates images to be identified. The lightweight feature extraction unit and the channel response recalibration unit sequentially complete feature extraction and effective feature enhancement. The identification output unit generates the equipment identification results, and the inspection data association mechanism completes the storage and display of the results. For continuous frame infrared images, the system executes the above process frame by frame according to the acquisition order and outputs the substation equipment identification results in each frame of the image.

[0093] S6.8: Configure the input image size, model inference accuracy, confidence threshold, non-maximum suppression threshold, and result output fields according to the computing power, storage conditions, and real-time requirements of the edge inspection platform.

[0094] The above configuration does not change the core recognition process that combines lightweight backbone feature extraction with channel response recalibration.

[0095] This invention provides a lightweight infrared image recognition method for substation equipment under end-side inspection. Compared with existing infrared image recognition methods for substation equipment, this invention can achieve lightweight model, enhanced effective infrared features, and stable recognition of multiple types of equipment targets under end-side computing power constraints. Its technical advantages are as follows: 1) In step 1, this invention establishes a unified preprocessing workflow for multi-source infrared inspection images. Addressing issues such as varying resolutions, inconsistent grayscale distributions, and significant local noise interference in images acquired by drones, inspection robots, fixed infrared monitoring terminals, and handheld infrared imaging devices, this invention converts the original infrared image into a suitable input image for the recognition network through size normalization, temperature grayscale mapping, noise suppression, and coordinate mapping parameter recording. This processing ensures consistency in the input format of infrared images from different sources and provides a foundation for subsequent bounding box coordinate inverse mapping, thereby improving the applicability and stability of the recognition workflow in actual inspection scenarios.

[0096] 2) In step 2, this invention constructs a lightweight backbone feature extraction network to replace the computationally intensive convolutional backbone in conventional target detection models. This network adopts a lightweight MobileNetV4-type backbone structure, extracting infrared image features step-by-step through initial convolutional layers and multi-level lightweight inverse residual feature units. It also utilizes channel-wise convolution and point-wise convolution to separate spatial texture modeling and channel information fusion processes. This reduces the number of parameters and multiply-accumulate operations while preserving the texture and structural semantic information of device edges, small targets, and other features in the infrared image, making the model more suitable for deployment on low-computing-power platforms such as UAV onboard computing units, inspection robots, and substation edge terminals.

[0097] 3) In step 3, this invention introduces a channel response recalibration module to effectively enhance the infrared feature map output by the lightweight backbone. This module obtains the global response information of each channel through global average pooling, then uses adaptive one-dimensional convolution to establish local channel dependencies, and generates channel weights to recalibrate the original feature map channel by channel. Compared with attention mechanisms using complex fully connected structures, the channel recalibration method adopted in this invention does not require channel dimensionality reduction, has lower computational overhead, and can enhance the target-related channels of the device under edge deployment conditions, suppressing the interference of background hotspots, gray-scale similar regions, and weakly correlated texture channels on the recognition results.

[0098] 4) In step 4, this invention utilizes a single-stage detection branch to perform target recognition on the recalibrated effective infrared feature map, directly outputting the equipment category, target confidence score, and bounding box coordinates. Through joint training of classification prediction, confidence score prediction, and bounding box regression, the model can simultaneously learn the category features and spatial location information of substation equipment. This design reduces the additional computational process caused by multi-stage candidate region generation, which is beneficial to improving the efficiency of infrared image recognition; at the same time, multi-scale feature input can take into account the recognition needs of equipment targets of different sizes, improving the detection stability of small-scale targets such as line hardware and insulators, as well as larger-scale targets such as surge arresters and high-voltage bushings.

[0099] 5) In step 5 of this invention, the candidate recognition results undergo confidence filtering, non-maximum suppression, and coordinate inversion processing, enabling the network output to be converted into data directly usable for inspection operations. Confidence filtering removes low-confidence candidate boxes, reducing misidentification caused by background areas; non-maximum suppression removes duplicate predicted boxes near the same target, improving the uniqueness of the output; and coordinate inversion restores the bounding boxes at the network input size to the original infrared image coordinates, ensuring the recognition results accurately correspond to the equipment locations in the original inspection image. This post-processing improves the reliability and engineering usability of the recognition results.

[0100] 6) In step 6, this invention integrates image acquisition, image preprocessing, lightweight feature extraction, channel response recalibration, detection output, and result management into an end-side inspection and recognition system. This system can be deployed in inspection robots, UAV onboard computing units, substation edge servers, or mobile inspection terminals, completing infrared image recognition and result output locally, reducing reliance on real-time transmission of raw images from remote servers. Recognition results can be associated and stored with inspection time, inspection point, equipment number, and raw image, or output to the inspection management platform in the form of visualized images or structured data, thereby improving the real-time performance and automation of infrared inspection of substation equipment. Attached Figure Description

[0101] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 A flowchart of the overall process for lightweight infrared image recognition of power equipment.

[0102] Figure 2 This is a schematic diagram of a lightweight backbone feature extraction network structure.

[0103] Figure 3 A schematic diagram of the channel response recalibration module structure.

[0104] Figure 4 This is a schematic diagram of a multi-scale, single-stage detection branch structure. Detailed Implementation

[0105] This embodiment provides a lightweight infrared image recognition method and system for substation equipment under end-side inspection. The method uses infrared images acquired during substation inspections as input, identifying targets including surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators. Under single-mode infrared image input conditions, the method utilizes a lightweight backbone feature extraction network to obtain multi-scale infrared features, enhances the relevant channel responses of equipment targets through a channel response recalibration module, and then outputs equipment category, target confidence, and bounding box coordinates through a multi-scale single-stage detection branch. The invention will be further described below with reference to the accompanying drawings and embodiments.

[0106] Specific steps are as follows Figure 1 As shown, it includes: Step 1: Acquire infrared images of the power equipment, preprocess the infrared images to obtain images to be identified that meet the input requirements of the lightweight backbone feature extraction network; Step 2: Construct a lightweight backbone feature extraction network. Input the image to be identified obtained in Step 1 into the lightweight backbone feature extraction network to obtain multiple infrared feature maps with different spatial resolutions. Step 3: Construct a channel response recalibration module at the output of the lightweight backbone feature extraction network to perform global channel statistics, local channel dependency modeling, and channel weight generation on the infrared feature map obtained in Step 2, and obtain the recalibrated effective infrared feature map. Step 4: Input the effective infrared feature map obtained in Step 3 into the detection branch to perform device category prediction, target confidence prediction and bounding box position regression to obtain candidate recognition results; Step 5: Perform confidence screening, non-maximum suppression, and coordinate inverse mapping on the candidate recognition results obtained in Step 4, and output the infrared image recognition results of the substation equipment.

[0107] Step 6: Construct a lightweight infrared image recognition system for power equipment. Deploy the recognition processes corresponding to Steps 1 to 5 on the end-side inspection platform to realize infrared image acquisition, preprocessing, lightweight recognition, result output, and inspection data management.

[0108] In step 1, acquiring and preprocessing the infrared image of the power equipment specifically includes: S1.1: As Figure 1 As shown, infrared images of substation equipment are collected using drone infrared cameras, inspection robot infrared imaging equipment, fixed infrared monitoring terminals, or handheld infrared imaging equipment. Each substation equipment infrared image contains at least one substation equipment target, including one or more of the following: surge arresters, high-voltage bushings, voltage transformers, line hardware, and insulators. The substation equipment category number, Chinese category, English category, and typical target characteristics can be set according to Table 1.

[0109]

[0110] S1.2: The acquired raw infrared images are formatted and read to obtain the raw infrared image matrix; if the raw infrared image is a temperature matrix, the temperature values ​​are mapped to a grayscale image; if the raw infrared image is already a pseudo-color or grayscale image, its grayscale intensity information or thermal radiation intensity information is extracted as subsequent recognition input. Specifically: Let the original infrared temperature matrix obtained be: (1); In equation (1), Represents spatial coordinates in the original infrared image The corresponding temperature value at that location; and These represent the pixel coordinates in the width and height directions of the image, respectively; and These represent the width and height of the original infrared image, respectively.

[0111] When the input data is a temperature matrix, the temperature values ​​are linearly mapped to a preset grayscale range based on the temperature value range. The calculation method is as follows: (2); In equation (2), Represents the pixel coordinates after temperature value mapping The grayscale value at that location; This indicates the number of gray levels. When using an 8-bit grayscale image... The corresponding grayscale value range is ; and These represent the lower and upper temperature limits used when performing grayscale mapping, respectively. They can be determined based on the lowest and highest temperatures of the current infrared image, or preset based on the effective temperature measurement range of the infrared imaging device. This indicates an extremely small positive number that prevents the denominator from being zero; This represents a cutoff function used to limit temperature values ​​that exceed a preset temperature range. Within the range.

[0112] When the input infrared image is a three-channel pseudo-color image, let the pixel values ​​of its red, green, and blue channels be respectively... , and Then it can be converted into a single-channel grayscale intensity: (3); In equation (3), , , Let these represent the grayscale conversion weights corresponding to the red, green, and blue channels, respectively, and satisfy the following conditions: (4); If the original input is already a single-channel grayscale infrared image, then directly set: (5); In equation (5), Represents the original grayscale infrared image in pixel coordinates The pixel intensity value at that location.

[0113] Step 1.3: Perform size normalization on the infrared images, scaling input images of different resolutions to a preset network input size. During scaling, record the original infrared image size, scaling ratio, and fill pixel range so that the bounding box coordinates output by subsequent detection branches can be mapped back to the original infrared image. Specifically: Let the original infrared image size be: (6); The default network input size is: (7); To avoid distortion of the device's geometric proportions caused by direct stretching, the image is scaled proportionally to its original aspect ratio. Represented as: (8); The scaled image dimensions are represented as follows: (9); (10); In the formula, and These represent the width and height of the image after proportional scaling, respectively. This represents the rounding operation that takes the nearest integer.

[0114] The total number of pixels that need to be filled in the width and height directions of the network input image are expressed as follows: (11); (12); When using a centrally symmetrical fill method, the fill amounts on the left and top are denoted as follows: and , is represented as: (13); In the actual pixel filling process, when or When the number is odd, the remaining pixels can be distributed to both sides of the image by rounding down and rounding up respectively.

[0115] For any pixel coordinate in the original image After being scaled and filled proportionally, its coordinates in the network input image are... Represented as: (14); (15); In the formula, Represents the pixel coordinates in the original infrared image; This represents the coordinates of the network input image after size normalization; Indicates the image scaling factor; and These represent the fill offsets in the width and height directions, respectively.

[0116] The parameter set is recorded in step 1.3: (16); This is used in step 5 to reverse map the predicted bounding box at the network input size to the original infrared image coordinates.

[0117] S1.4: Perform grayscale normalization processing on the infrared image to normalize the image pixel values ​​or temperature mapping values ​​to a uniform range, as follows: Let the infrared image after temperature grayscale mapping or grayscale reading be... Its grayscale value range is Then the normalized pixel value Represented as: (17); In equation (17), Represents the normalized pixel coordinates The pixel value at that location has the following range: (18); and These represent the minimum and maximum gray values ​​of the image before normalization, respectively. This represents the smallest positive number that prevents the denominator from being zero.

[0118] When the input image is represented using 8-bit grayscale, and the grayscale value range is fixed. When the above formula simplifies to: (19); Through the above normalization process, infrared images from different sources can have a uniform range of numerical values ​​when entering the recognition network, thereby reducing the impact of the absolute amplitude difference of input pixels on the subsequent feature extraction process.

[0119] S1.5: Perform noise suppression and edge preservation processing on the infrared image: To address random noise, localized hotspots, and weakly textured areas in infrared images, median filtering, bilateral filtering, or lightweight smoothing techniques can be used to reduce noise interference while preserving as much of the substation's edge contours and target structural information as possible. Specifically, this includes: Let the input infrared image be The infrared image after noise suppression processing is ,in: This represents the pixel coordinates in an infrared image.

[0120] ①: When using median filtering, the filtered pixel values Represented as: (20); In equation (20): Represented by pixels A local neighborhood window centered on the target; This represents the pixel coordinates within the local neighborhood window; This represents taking the median of pixel grayscale values ​​within a local neighborhood window. This method can suppress isolated noise points and localized anomalous grayscale responses in infrared images. ②: When using bilateral filtering, the filtered pixel values As shown in formula (21): (twenty one); In equation (21): Represents the normalization coefficient; This represents the spatial distance weighting function. This represents the gray-level similarity weighting function; Represents the two-dimensional Euclidean distance. Indicates input infrared image In neighboring pixel coordinates The grayscale value or the pixel intensity value after temperature mapping at that location; Represents the center pixel coordinates The grayscale value or pixel intensity value at that location.

[0121] Among them, the normalization coefficient As shown in formula (22): (twenty two); The spatial distance weight function and the gray-level similarity weight function are respectively expressed as: (twenty three); (twenty four); In the formula, Represented by natural constant An exponential function with base 0; This represents the spatial domain standard deviation, used to control the impact of the spatial distance between neighboring pixels on the filter weights; This represents the standard deviation in the grayscale domain, used to control the impact of grayscale differences between neighboring pixels on the filtering weights. This method allows for the smoothing of infrared image noise while preserving the edge contour information of the substation target. ③: When using the lightweight smoothing method, the filtered pixel values As shown in equation (25): (25); In equation (25): Represents a local neighborhood window The number of pixels within the image. This method has low computational complexity and is suitable for fast image preprocessing in scenarios with limited computing power, such as drones, inspection robots, or edge computing terminals.

[0122] The above processing can reduce the impact of random noise, local abnormal gray points and weak background interference in infrared images on subsequent feature extraction, while preserving as much as possible the thermal edge, main outline and local structural information of the substation target.

[0123] S1.6: During the training phase, the device category labels and bounding box coordinates in the annotation file are synchronously transformed to correspond to the preprocessed network input image; specifically as follows: During the inference phase, only the image scaling and padding parameters are retained for coordinate inverse mapping in step 5.

[0124] Let the first part of the original infrared image be... The category label for each target is Its actual labeled bounding box is represented as: (26); In equation (26), and They represent the first The pixel coordinates of the top left and bottom right corners of the target bounding box in the original infrared image.

[0125] Based on the scaling ratio obtained in step 1.3 and fill offset , The coordinates of the bounding box after synchronous transformation are: (27); (28); (29); (30); Accordingly, the bounding box after synchronous transformation is represented as: (31); The equipment category label remains unchanged during geometric transformation: (32); When the training annotations use normalized center point coordinates and width / height, further calculations can be performed: (33); (34); (35); (36); In the formula, They represent the first The normalized coordinates of the center point of each target bounding box relative to the width and height of the network input image; and These represent the normalized bounding box width and height, respectively; and These represent the width and height of the input image to the network, respectively.

[0126] Step 2, which involves constructing a lightweight backbone feature extraction network, specifically includes: S2.1: As Figure 2 As shown, a lightweight backbone feature extraction network for infrared images is constructed based on a single-stage target detection framework. The lightweight backbone feature extraction network consists of an initial convolutional layer and multiple levels of lightweight inverse residual feature units, which are used to extract multi-scale features from the input infrared image, ranging from shallow edge texture to deep semantic information.

[0127] S2.2: Input the image to be identified obtained in step 1 into the initial convolutional layer to obtain the initial infrared feature map. The initial convolutional layer is used to complete the mapping of the infrared image from pixel space to feature space, and to initially extract information on the thermal edges, local textures, and regional grayscale differences of the power equipment. Specifically: Let the network input infrared image obtained in step 1 be: (37); The initial infrared feature map is obtained through the initial convolutional layer. , is represented as: (38); Furthermore, the first The output channels are located in space. The eigenvalues ​​at that location are represented as: (39); In the above formula, Represents the input infrared image tensor; Indicates the number of input channels; and These represent the height and width of the network input image, respectively. This represents the infrared feature map output by the initial convolutional layer; Indicates the output channel number; Indicates the input channel number; Indicates the initial convolution kernel size; and Indicates the position index inside the convolution kernel; This represents the initial convolutional layer kernel parameters; Indicates the bias parameter; Indicates the convolution stride; Indicates the fill size; This represents the convolution operation; This indicates a batch normalization operation; This represents a non-linear activation function.

[0128] Through this convolutional mapping, the raw pixel intensity of the input infrared image is converted into a multi-channel feature response, enabling the subsequent network to progressively extract the thermal edge, contour, and local structural information of the device in the feature space.

[0129] S2.3: The initial infrared feature map is input into a multi-level lightweight inverse residual feature unit to sequentially obtain infrared feature maps with different spatial resolutions. The shallow high-resolution feature map is used to preserve device edges and small target textures, the mid-level semantic feature map is used to express the device's structural morphology, and the deep low-resolution feature map is used to provide semantic information about the device category. Details are as follows: Let the initial convolutional layer output be , No. The lightweight feature extraction unit is denoted as The multi-level feature extraction process can then be represented as: (40); In equation (40), Indicates the number of levels in a lightweight feature extraction network; and They represent the first The input and output feature maps of the level feature extraction unit; Indicates the first Lightweight feature transformation.

[0130] Three scale features were selected from different network depths and denoted as follows: (41); (42); (43); In the formula, , and These represent infrared feature maps at three scales with step sizes of 8, 16, and 32, respectively. , and These represent the number of channels in the corresponding scale feature map; and This indicates the height and width of the input image to the network.

[0131] When the network input size is At that time, the corresponding spatial dimensions are as follows: (44); Its multi-scale output modes of P3 / 8, P4 / 16, and P5 / 32 correspond to the three detection scale settings of the YOLOv11 detection structure.

[0132] S2.4: The lightweight inverse residual feature unit comprises a channel expansion layer, a spatial modeling layer, a channel projection layer, and a residual connection layer. The channel expansion layer uses 1×1 convolutions to expand the input features, enhancing the feature representation dimension. The spatial modeling layer uses depthwise convolutions or depthwise separable convolutions to extract local infrared spatial features. The channel projection layer uses 1×1 convolutions to compress the number of channels, reducing computational cost. When the input and output feature sizes and number of channels are consistent, the residual connection layer adds the input features to the projected features to enhance gradient propagation and reduce information loss caused by lightweight compression. Details are as follows: Let the input features of the lightweight reciprocal residual feature unit be: (45); First, optional channel-wise spatial feature extraction is performed: (46); Then adopt Channel expansion via pointwise convolution: (47); The expanded number of channels is expressed as: (48); In equation (48), Indicates the channel spread factor; This indicates the number of feature channels after expansion.

[0133] Then perform channel-wise spatial convolution: (49); Then through Pointwise convolution completes channel projection: (50); When the input and output feature spaces have the same size and the same number of channels, residual connection is used: (51); otherwise: (52); In the above formula, This represents channel-wise convolution; express Pointwise convolution; and These represent the kernel size and stride of the preceding channel-wise convolution, respectively. and These represent the kernel size and stride of the channel-wise spatial convolution after channel expansion, respectively. This represents the number of channels for the input feature; This indicates the number of feature channels after channel expansion; This represents the feature obtained after channel projection; This represents the final output feature of the lightweight inverse residual feature unit.

[0134] S2.5: The lightweight backbone feature extraction network adopts a MobileNetV4-type backbone structure and selects a convolutional small model configuration as the backbone parameter configuration. By decomposing standard convolution operations through channel-wise spatial convolution and point-wise channel-wise convolution, infrared spatial texture extraction and cross-channel information fusion are performed separately, thereby reducing the number of parameters and multiplication-accumulation operations in the feature extraction stage. Specifically: In standard convolution, the first The eigenvalues ​​of each output channel are represented as follows: (53); In equation (53), Represents the first... One channel; The first part represents the output feature map. One channel; Indicates the standard convolution kernel parameters; Indicates the number of input channels; This represents the side length of the convolution kernel.

[0135] When using depthwise separable convolution, spatial convolution is first performed on each input channel separately: (54); Then adopt Pointwise convolution completes cross-channel information fusion: (55); In the formula, Indicates the first Intermediate features obtained by channel-wise spatial convolution of each input channel; Indicates the first The depthwise convolution kernel parameters corresponding to each channel; Indicates by the first The intermediate channel is mapped to the first... Each output channel Pointwise convolution parameters.

[0136] Therefore, the standard convolution parameter count is: (56); The number of parameters for depthwise separable convolution is: (57); S2.6: Let the number of input feature map channels be... The number of channels in the output feature map is The kernel size is The feature map space size is Standard convolution computational cost Depthwise separable convolution The computational load is expressed as follows: (58); (59); In the formula, and These represent the height and width of the output feature map, respectively; and These represent the number of channels in the input feature map and the output feature map, respectively. Indicates the side length of the convolution kernel; Indicates the computational cost of standard convolution; The computational cost of depthwise separable convolution is expressed as follows. Equation (58) represents the computational cost of standard convolution, and Equation (59) represents the computational cost of depthwise separable convolution. In Equation (59), the first term represents the computational cost of channel-wise spatial convolution, and the second term represents... The computational cost of cross-channel information fusion using pointwise convolution.

[0137] As can be seen from the above formula, under the same input and output channel number and feature map size, depthwise separable convolution can reduce the number of convolutional layer parameters and multiply-accumulate operations by separating the extraction of spatial features and the fusion of cross-channel information. This reduces the computational complexity of the lightweight backbone network, making the present invention more suitable for deployment on edge inspection equipment with limited computing power, such as UAV onboard computing units, inspection robots and station edge terminals.

[0138] S2.7: To ensure that the output features of the backbone network can be adapted to subsequent detection branches, the infrared feature maps of different scales output by the multi-level lightweight inverse residual feature units are numbered according to scale, and are designated as P3, P4 and P5. Among them, P3 retains higher resolution and more detailed information, and is mainly used for small-scale substation equipment target recognition; P4 takes into account both spatial information and semantic expression; P5 contains stronger semantic information and is used for larger-scale equipment targets and overall structure recognition.

[0139] Step 3 specifically includes: S3.1: Input the infrared feature map of at least one scale obtained in step 2 into the channel response recalibration module. The channel response recalibration module is located between the output of the lightweight backbone feature extraction network and the input of the detection branch. It adaptively evaluates the importance of different channels in the infrared feature map and enhances the device target-related feature response without significantly increasing the number of parameters and computational load. The structure of the channel response recalibration module is as follows: Figure 3 As shown.

[0140] S3.2: As Figure 3 As shown, for any scale of input infrared feature map ,in: Indicates the number of channels. and These represent the height and width of the feature map, respectively. The channel response recalibration module processes the data using a parallel approach that preserves the original feature bypass and generates channel weight branches. Input feature map On the one hand, it is directly passed along the bypass branch to preserve the original spatial structure information; on the other hand, it is input into the global average pooling layer (GAP) to extract the global response statistics of each channel. For the Each channel undergoes global average pooling, and its channel descriptor value Represented as: (60); In equation (60), Indicates the input infrared feature map The Middle The spatial location of each channel eigenvalues ​​at This represents the spatial position index along the height direction of the feature map, with a value range of [value range missing]. ; This represents the spatial position index along the width direction of the feature map, with a value range of [value range missing]. ; Indicates the first The channel description values ​​obtained after global average pooling of each channel; S3.3: Obtaining the channel description vector Then, based on the number of channels in the input feature map Adaptive determination of one-dimensional convolution kernel size .like Figure 3 middle As shown, the kernel size With the number of channels Correspondingly, infrared feature maps of different channel sizes can employ matching local channel interaction ranges. The adaptive convolutional kernel size can be expressed as: (61); In equation (61), and Preset adjustment parameters; This indicates taking the closest odd number. This method allows the channel interaction range to adaptively adjust with the number of feature map channels without artificially fixing the convolution kernel size.

[0141] S3.4: The one-dimensional convolution kernel size determined using S3.3 , channel description vector Perform one-dimensional convolution operations to establish local dependencies between adjacent channels and obtain channel interaction response vectors. The calculation process can be expressed as follows: (62); In equation (62), Indicates the kernel size as One-dimensional convolution operation. Compared with attention structures that use fully connected layers for channel dimensionality reduction and enhancement, this invention performs local interaction modeling directly in the channel dimension through one-dimensional convolution, without changing the channel dimension. This avoids the loss of infrared detail information caused by channel dimensionality reduction and reduces the number of additional parameters.

[0142] S3.5: The channel interaction response vector obtained in S3.4 Inputting the Sigmoid activation function yields the channel weight vector. The calculation process can be expressed as follows: (63); In equation (63), This represents the Sigmoid activation function; This represents the generated channel weight vector.

[0143] S3.6: The channel weight vector obtained in S3.5 Compared with the original infrared feature map preserved by the bypass in S3.2 Perform channel-by-channel multiplication to obtain the recalibrated effective infrared feature map. Its calculation process can be expressed as equation (64): (64); Or it can be expressed as equation (65): (65); In the formula, Indicates the first Output characteristics after recalibration of each channel; This indicates a channel-by-channel multiplication operation. Through this processing, the responses of channels related to device thermal edges, main body contours, and structural textures are enhanced, while the responses of background hotspots, reflective areas, and weakly related texture channels are suppressed, thereby improving the ability to distinguish infrared device targets from background areas.

[0144] S3.7: For infrared feature maps of different scales, execute S3.2~S3.6 respectively to independently generate channel weights for the corresponding scales and complete feature recalibration. When the input feature map is a shallow, high-resolution feature, the channel response recalibration module focuses on enhancing device edges, small target textures, and local thermal structure information; when the input feature map is a mid-level feature, it focuses on enhancing device morphology and local structural information; when the input feature map is a deep, low-resolution feature, it focuses on enhancing semantically relevant channels for device categories. This achieves differentiated enhancement of infrared features at different scales.

[0145] S3.8: Output the recalibrated infrared effective feature maps at each scale to the subsequent detection branch. Through the channel response recalibration module, the detection branch can prioritize the use of effective channel information related to the substation equipment target when performing equipment category prediction, target confidence prediction, and bounding box regression, thereby reducing the interference of invalid background responses on the recognition results and improving the stability of infrared equipment target recognition under end-side inspection conditions.

[0146] Step 4 specifically includes: S4.1: As Figure 4 As shown, the effective infrared feature maps of multiple scales obtained in S3 are input into a single-stage detection branch, which specifically includes a classification prediction sub-branch, a target confidence prediction sub-branch, and a bounding box regression sub-branch. The classification prediction sub-branch is used to determine the equipment category to which the candidate target belongs, the target confidence prediction sub-branch is used to determine whether there is a substation equipment target in the candidate region, and the bounding box regression sub-branch is used to predict the location of the equipment target in the image.

[0147] The classification prediction sub-branch outputs the probability value of each candidate target belonging to the categories of surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators. The target confidence prediction sub-branch outputs the probability that the candidate region contains substation equipment targets. The bounding box regression sub-branch outputs the coordinates of the bounding box center point, width, and height. Details are as follows: For the Each scale feature map The output of the detection branch is represented as: (66); In equation (66), This indicates three detection scales; Indicates the first Each scale corresponds to a detection head; This represents the bounding box regression output; This represents the target confidence prediction output; This indicates the category prediction output.

[0148] For feature map location First The probability of a device class is expressed as follows: (67); in, (68); In the formula, Indicates the output of the detection head. Class prediction value; Indicates the probability of the corresponding category; ,in This indicates the total number of equipment categories.

[0149] The predicted device category is: (69); The corresponding target confidence level is defined as: (70); In the formula, Indicates the location of the feature map The predicted confidence level of the target at the location; This represents the Sigmoid activation function; This indicates the confidence level that the corresponding candidate region contains the target of the power equipment.

[0150] Suppose that the bounding box regression yields the predicted bounding box: (71); This can then be further converted into a center point and width / height format: (72); (73); (74); (75); In the formula, and These represent the coordinates of the top-left and bottom-right corners of the predicted bounding box, respectively. Indicates the coordinates of the center point of the bounding box; and These represent the width and height of the bounding box, respectively.

[0151] Therefore, the first The candidate identification results can be uniformly represented as: (76); In equation (76), , Indicates the first One candidate identification result; Indicates the predicted device category; Indicates the confidence level of the target; Indicates the coordinates of the predicted bounding box center point; and These represent the width and height of the predicted bounding box, respectively.

[0152] For each predicted target, the detection branch outputs its class probability, target confidence, and bounding box coordinates. Each scale feature map corresponds to three types of outputs: classification prediction, confidence prediction, and bounding box regression.

[0153] S4.2: During the training phase, the lightweight backbone feature extraction network, channel response recalibration module, and detection branch are jointly trained based on the category labels and bounding box labels in the infrared image dataset of substation equipment.

[0154] Training losses include classification loss, object confidence loss, and bounding box regression loss: The detection branch employs a joint loss function to optimize classification prediction, target confidence prediction, and bounding box regression. Specifically, the classification loss constrains the consistency between the predicted and ground truth classes, the target confidence loss determines whether the candidate region contains the substation target, and the bounding box regression loss constrains the degree of overlap between the predicted and ground truth bounding boxes. Let the predicted bounding box be The actual labeled bounding box is The intersection and union ratio of the two Represented as: (77); In equation (77), This represents the intersection area of ​​the predicted bounding box and the ground truth bounding box. This represents the area of ​​the union of the two. Based on the intersection-union ratio, the bounding box regression loss... As shown in equation (78): (78); Total loss function for detection branches As shown in equation (79): (79); In equation (79), Represents classification loss, Indicates the target confidence loss. This represents the bounding box regression loss; , , These represent the weighting coefficients of each loss term. Using the joint loss function described above, the detection branch can improve the matching degree between the predicted bounding box and the real target region while simultaneously identifying the device category. S4.3: During the inference phase, the detection branch generates candidate recognition boxes based on the infrared effective feature map output in step 3, and obtains the device category, confidence level and bounding box coordinates of each candidate target.

[0155] Step 5 specifically includes: S5.1: Read the candidate recognition results obtained in 4.3. The candidate recognition results include the device category, target confidence level, and bounding box coordinates of the candidate target. Establish a candidate box set for all candidate recognition results in the same infrared image, denoted as: (80); In equation (80), Represents the set of candidate recognition results. Indicates the first One candidate identification result, This indicates the number of candidate identification results.

[0156] S5.2: Confidence screening of candidate identification results: Let the first The target confidence level for each candidate identification result is: The preset reliability threshold is ,when When, delete the candidate recognition result; when If a candidate box is found to be falsely identified, it is retained for subsequent processing. This method removes low-confidence candidate boxes, reducing false identifications caused by background areas, hotspot areas, and weak response areas.

[0157] S5.3: The candidate identification results after confidence level screening are grouped according to equipment category. For different categories such as surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators, candidate box sorting and duplicate box suppression are performed separately to avoid mutual interference between candidate box screenings of different categories. Details are as follows: Let the set of candidate identification results after confidence level filtering be: (81); Among them, the The candidate identification results are represented as follows: (82); In the formula, Indicates the first Predicted bounding boxes of candidate targets; Indicates the predicted device category; Indicates the confidence level of the target; This represents the total number of candidate boxes retained after confidence level filtering.

[0158] For the For each device class, create a corresponding set of candidate frames: (83); In the formula, This represents the total number of categories of power equipment; in this embodiment... .

[0159] For sets Candidate results in the data are ranked according to confidence level. Sort in descending order. Let the sorted index sequence be: (84); The sorted candidate box set is then represented as: (85); And satisfy: (86); In the formula, Indicates the first The number of candidate boxes corresponding to each type of device; This indicates an operation that returns the index of the sorted element.

[0160] S5.4: Perform non-maximum suppression processing on candidate bounding boxes within the same device category. Specifically: First, the candidate boxes are sorted from highest to lowest target confidence level, and the candidate box with the highest confidence level is selected as the current retained box; the details are as follows: Let the first The current set of candidate boxes to be processed is ,in The first nonmaximum suppression represents the... In the next iteration, the index of the candidate box with the highest confidence is represented as: (87); The corresponding candidate results Add to the final reserved set : (88); In the formula, Indicates the first The index of the candidate box with the highest target confidence in the next iteration; Indicates the first The target confidence level of each candidate box; Indicates the first The first iteration has already retained the... A collection of candidate boxes.

[0161] Then, the intersection-union ratio (IUR) between the remaining candidate boxes and the currently retained box is calculated. If the IUR is greater than a preset threshold, it is considered that the two correspond to the same substation equipment target, and the candidate boxes with lower confidence are judged as duplicate candidate boxes and deleted; specifically as follows: Let the currently retained box be: (89); The candidate boxes to be compared are: (90); The width and height of the intersection region are represented as follows: (91); (92); The area of ​​intersection is: (93); The areas of the two candidate boxes are as follows: (94); (95); The intersection, union, and ratio of the two can be expressed as: (96); Let the nonmaximum suppression threshold be... ,when: (97); and: (98); Then the candidate box Determined as the currently retained frame The corresponding duplicate prediction boxes are then removed from the set to be processed.

[0162] The candidate box set after one iteration is represented as: (99); Repeat the above process until: (100); at this time, That is, the first The set of candidate recognition results retained by the device after nonmaximum suppression.

[0163] Repeat the above process until all candidate boxes in this category have been processed, thereby reducing the number of times the same target on the same device is selected multiple times.

[0164] S5.5: Perform coordinate inverse mapping on the bounding boxes retained after nonmaximum suppression: Based on the image scaling ratio, fill size, and original image size recorded in step 1, the bounding box coordinates under the network input size are converted back to the original infrared image coordinates to obtain the actual position of the device target in the original image. Specifically: Let the coordinates of the predicted bounding box retained after nonmaximum suppression in the network input image be: (101); Based on the image scaling ratio recorded in step 1.3 And the fill offset in the horizontal and vertical directions. and The predicted bounding box is then mapped inversely to the original infrared image coordinate system, as follows: (102); In the formula, This represents the predicted bounding box in the coordinate system of the network input image; This represents the bounding box after inversion onto the original infrared image; and These represent the coordinates of the top-left and bottom-right corners of the predicted bounding box in the network input image, respectively. This represents the scaling factor of the image recorded in step 1.3; and These represent the fill offsets of the image in the horizontal and vertical directions, respectively.

[0165] S5.6: Output the final infrared image recognition results of the substation equipment: The recognition results include at least the device category, bounding box coordinates, target confidence score, and image number; Depending on the actual inspection needs, it may also include the collection time, inspection point, equipment number, collection terminal number, and identification result storage path.

[0166] S5.7: Associate the final recognition result with the original infrared image. For scenarios requiring visualization, the device category name, target confidence score, and bounding box position can be overlaid onto the original infrared image to form a visual recognition map; as detailed below: For scenarios requiring data retrieval, the recognition results can be output to the inspection management platform in the form of structured data.

[0167] Let the first The final identification result is: (103); in: (104); Indicates the final bounding box; Indicates the equipment category; Indicates the confidence level of the target.

[0168] Definition of the first The set of boundary pixels of each bounding box is: (105); For all final bounding boxes, define a visual boundary mask: (106); The image after the bounding boxes are superimposed can be represented as: (107); In equation (107), Represents the original infrared image; This represents a visualized image after the recognition results have been overlaid. Represents the bounding box mask; Indicates the relationship with the first The preset display value or display color corresponding to this type of device.

[0169] No. The visual text labels corresponding to each target are represented as follows: (108); In equation (108), Indicates based on category number The device category name obtained from the query; This indicates the target confidence level after formatting according to a preset number of decimal places; This indicates a string concatenation operation.

[0170] Text Label It can be set near the top left corner of the bounding box: (109); This results in a visual recognition outcome that includes the device category name, target confidence level, and spatial location.

[0171] S5.8: The data fields of the final recognition result can be set according to Table 2. Table 2 may include field names, field meanings, and data types, which are used to unify the storage format and interface output format of the recognition results.

[0172]

[0173] It also includes step 6: Constructing a lightweight infrared image recognition system for power equipment, deploying the recognition processes corresponding to steps 1 to 5 on the end-side inspection platform to realize infrared image acquisition, preprocessing, lightweight recognition, result output and inspection data management.

[0174] In step 6, a lightweight infrared image recognition system for substation equipment is constructed and deployed on the end-side inspection platform, specifically including: S6.1: As Figure 1 As shown in the end-side deployment and application section, an infrared image input interface is set in the end-side inspection platform. The infrared image input interface is connected to the UAV infrared camera, the inspection robot infrared imaging equipment, the fixed infrared monitoring terminal or the handheld infrared imaging equipment to receive infrared images collected during the substation inspection process and send the infrared images into the subsequent image processing flow.

[0175] S6.2: Establish a preprocessing procedure for the input infrared image. The preprocessing procedure includes image size normalization, temperature grayscale mapping, noise suppression, pixel normalization, and coordinate mapping parameter recording. After preprocessing, the original infrared image is converted into a recognition image that meets the input requirements of the lightweight recognition network, while retaining the image scaling ratio and padding information so that the subsequent recognition results can be mapped back to the original image coordinates.

[0176] S6.3: The image to be identified is fed into the lightweight feature extraction unit, which performs the backbone feature extraction process in step 2 to obtain infrared feature maps with different spatial resolutions. This unit reduces the number of parameters and computational load during edge-side inference through a lightweight backbone network, providing a multi-scale feature foundation for subsequent device target recognition.

[0177] S6.4: After the lightweight feature extraction unit, a channel response recalibration unit is set up to perform the channel weight calculation and feature recalibration processing in step 3 on the multi-scale infrared feature map. Through this processing, the channels related to the device's thermal edge, main body contour, and structural texture are enhanced, while the channel responses corresponding to background hot spots and weakly correlated areas are suppressed, thereby obtaining an effective infrared feature map.

[0178] S6.5: The effective infrared feature map is input into the recognition output unit, which performs candidate box generation, confidence level filtering, non-maximum suppression, and coordinate inverse mapping processing, and outputs the infrared image recognition result of the substation equipment. The recognition result includes equipment category, equipment location, and target confidence level.

[0179] S6.6: Store the equipment category, bounding box coordinates, target confidence score, inspection task number, acquisition time, inspection point, equipment number, and original infrared image. For scenarios requiring manual verification, the recognition box, category name, and confidence score can be overlaid on the original infrared image to generate a visual recognition result; for scenarios requiring platform access, the recognition result can be converted into structured data and output to the inspection management platform.

[0180] S6.7: During the operation of the end-side inspection platform, the infrared image input interface receives inspection images in real time. The preprocessing process generates images to be identified. The lightweight feature extraction unit and the channel response recalibration unit sequentially complete feature extraction and effective feature enhancement. The identification output unit generates the equipment identification results, and the inspection data association mechanism completes the storage and display of the results. For continuous frame infrared images, the system executes the above process frame by frame according to the acquisition order and outputs the substation equipment identification results in each frame of the image.

[0181] S6.8: Configure the input image size, model inference accuracy, confidence threshold, non-maximum suppression threshold, and result output fields according to the computing power, storage conditions, and real-time requirements of the edge inspection platform.

[0182] The above configuration does not change the core recognition process that combines lightweight backbone feature extraction with channel response recalibration.

[0183] To verify the feasibility of the method in this embodiment, an experiment was conducted using an infrared image dataset of substation equipment. This dataset consisted of infrared images collected during substation equipment operation and inspection, encompassing five types of substation equipment targets: surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators. The experimental images were uniformly adjusted to the network input size and divided into training, validation, and test sets.

[0184] To verify the detection performance and deployment adaptability of the method of the present invention, YOLOv11n was used as the basic control model, and experiments were carried out under the same training configuration, data partitioning method and testing conditions. The experimental results are shown in Table 3.

[0185]

[0186] The results show that the mAP50 of the method of this invention is 90.9%, which is higher than that of models such as SSD, Faster R-CNN, YOLOv5n, and YOLOv6n, indicating that it can detect substation targets more completely and improve the target localization effect. In terms of accuracy, the method of this invention reaches 98.6%, higher than all the comparative methods, indicating that the number of false detections is effectively controlled. The parameter count of the method of this invention is 1.26M, significantly smaller than traditional detection models and other YOLO series comparative models, thus making it more suitable for edge platforms. Regarding inference speed, the method of this invention achieves 101.2 FPS under the same test conditions, higher than the original YOLOv11n model, indicating that it has good real-time detection capabilities.

[0187] Experiments show that this embodiment can maintain good recognition accuracy and real-time reasoning ability while reducing model size and computational complexity, and is suitable for application scenarios such as UAV inspection of substations, robot inspection and real-time recognition of edge terminals.

Claims

1. A lightweight infrared image recognition method for substation equipment for end-side inspection, characterized in that... Includes the following steps: Step 1: Acquire infrared images of the power equipment, preprocess the infrared images to obtain images to be identified that meet the input requirements of the lightweight backbone feature extraction network; Step 2: Construct a lightweight backbone feature extraction network. Input the image to be identified obtained in Step 1 into the lightweight backbone feature extraction network to obtain multiple infrared feature maps with different spatial resolutions. Step 3: Construct a channel response recalibration module at the output of the lightweight backbone feature extraction network to perform global channel statistics, local channel dependency modeling, and channel weight generation on the infrared feature map obtained in Step 2, and obtain the recalibrated effective infrared feature map. Step 4: Input the effective infrared feature map obtained in Step 3 into the detection branch to perform device category prediction, target confidence prediction and bounding box position regression to obtain candidate recognition results; Step 5: Perform confidence screening, non-maximum suppression, and coordinate inverse mapping on the candidate recognition results obtained in Step 4, and output the infrared image recognition results of the substation equipment.

2. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 1, characterized in that: In step 1, acquiring and preprocessing the infrared image of the power equipment specifically includes: S1.1: Collect infrared images of substation equipment during inspection. The infrared images of substation equipment during inspection contain at least one type of substation equipment target, including one or more of surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators; the category number, Chinese category, English category, and typical target characteristics of the substation equipment. S1.2: The acquired raw infrared images are formatted and read to obtain the raw infrared image matrix; if the raw infrared image is a temperature matrix, the temperature values ​​are mapped to a grayscale image; if the raw infrared image is already a pseudo-color or grayscale image, its grayscale intensity information or thermal radiation intensity information is extracted as subsequent recognition input; specifically as follows: Let the original infrared temperature matrix obtained be: （1）； In equation (1), Represents spatial coordinates in the original infrared image The corresponding temperature value at that location; and These represent the pixel coordinates in the width and height directions of the image, respectively; and These represent the height and width of the original infrared image, respectively; express OK The space of real matrices in the column; When the input data is a temperature matrix, the temperature values ​​are linearly mapped to a preset grayscale range based on the temperature value range. The calculation method is as follows: （2）； In equation (2), Represents the pixel coordinates after temperature value mapping The grayscale value at that location; This indicates the number of gray levels. When using an 8-bit grayscale image... ; and These represent the lower and upper temperature limits used when performing grayscale mapping, respectively, which are determined based on the lowest and highest temperatures of the current infrared image, or preset based on the effective temperature measurement range of the infrared imaging device. This indicates a very small positive number that prevents the denominator from being zero; This represents a cutoff function used to limit temperature values ​​outside a preset temperature range. Within the range; When the input infrared image is a three-channel pseudo-color image, let the pixel values ​​of its red, green, and blue channels be respectively... , and Then it can be converted into a single-channel grayscale intensity: （3）； In equation (3), , , Let these represent the grayscale conversion weights corresponding to the red, green, and blue channels, respectively, and satisfy the following conditions: （4）； If the original input is already a single-channel grayscale infrared image, then directly set: （5）； In equation (5), Represents the original grayscale infrared image in pixel coordinates Pixel intensity value at; Step 1.3: Perform size normalization on the infrared images, scaling input images of different resolutions to a preset network input size; during scaling, record the original infrared image size, scaling ratio, and fill pixel range so that the bounding box coordinates output by subsequent detection branches can be mapped back to the original infrared image; details are as follows: Let the original infrared image size be: （6）； The default network input size is: （7）； in, Enter the height for the network. Enter the network width; To avoid distortion of the device's geometric proportions caused by direct stretching, the image is scaled proportionally to its original aspect ratio. Represented as: （8）； This is the scaling factor for the image; This indicates the operation of taking the minimum value; The scaled image dimensions are represented as follows: （9）； （10）； In the formula, and These represent the width and height of the image after proportional scaling, respectively. This indicates a rounding operation that takes the integer closest to the given number. The total number of pixels that need to be filled in the width and height directions of the network input image are expressed as follows: （11）； （12）； In the formula, This indicates the total number of pixels that need to be filled in the width direction of the network input image; This represents the total number of pixels that need to be filled in the height direction of the network input image; When using a centrally symmetrical fill method, the fill amounts on the left and top are denoted as follows: and , represented as: （13）； In the actual pixel filling process, when or When the number is odd, the remaining pixels are distributed to both sides of the image by rounding down and rounding up respectively; For any pixel coordinate in the original image After being scaled and filled proportionally, its coordinates in the network input image are... Represented as: （14）； （15）； In the formula, Represents the pixel coordinates in the original infrared image; This represents the coordinates of the network input image after size normalization; Indicates the image scaling factor; and These represent the fill offsets in the width and height directions, respectively. The parameter set is recorded in step 1.3: （16）； This represents the set of mapping parameters, which records the original size, network input size, scaling ratio, and padding offset. These parameters are used in step 5 to reverse map the predicted bounding box under the network input size to the original infrared image coordinates. S1.4: Perform grayscale normalization processing on the infrared image to normalize the image pixel values ​​or temperature mapping values ​​to a uniform range, as follows: Let the infrared image after temperature grayscale mapping or grayscale reading be... Its grayscale value range is Then the normalized pixel value Represented as: （17）； In equation (17), , These represent the minimum and maximum gray values ​​of the image before normalization, respectively. Represents the normalized pixel coordinates The pixel value at that location has the following range: （18）； and These represent the minimum and maximum gray values ​​of the image before normalization, respectively. This indicates an extremely small positive number that prevents the denominator from being zero; When the input image is represented using 8-bit grayscale, and the grayscale value range is fixed. When the above formula simplifies to: （19）； S1.5: Perform noise suppression and edge preservation processing on the infrared image: To address random noise, local hot spots, and weak texture areas in infrared images, median filtering, bilateral filtering, or lightweight smoothing methods are employed to reduce noise interference while preserving the edge contours of the substation and the target structure information.

3. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 2, characterized in that: S1.5 specifically includes: Let the input infrared image be The infrared image after noise suppression processing is ,in: Represents the pixel coordinates in an infrared image; ①: When using median filtering, the filtered pixel values Represented as: （20）； In equation (20): Represented by pixels A local neighborhood window centered on the target; This represents the pixel coordinates within the local neighborhood window; This indicates that the median grayscale value of the pixels within the local neighborhood window is taken; This represents the pixel value after median filtering; Represents neighboring pixels The grayscale value at that location; ②: When using bilateral filtering, the filtered pixel values As shown in formula (21): （21）； In equation (21): This represents the pixel value after bilateral filtering; Represents the normalization coefficient; This represents the spatial distance weighting function. This represents the gray-level similarity weighting function; Represents the two-dimensional Euclidean distance. Represents the center pixel coordinates The grayscale value or pixel intensity value at that location; Among them, the normalization coefficient As shown in formula (22): （22）； The spatial distance weight function and the gray-level similarity weight function are respectively expressed as: （23）； （24）； In the formula, Represented by natural constant An exponential function with base 0; This represents the spatial domain standard deviation, used to control the impact of the spatial distance between neighboring pixels on the filter weights; This represents the standard deviation of the grayscale domain, used to control the impact of grayscale differences between neighboring pixels on the filter weights. This represents the squared spatial distance between the center pixel and its neighboring pixels; This represents the squared grayscale difference between the center pixel and its neighboring pixels; ③: When using the lightweight smoothing method, the filtered pixel values As shown in equation (25): （25）； In equation (25): Represents a local neighborhood window The number of pixels within; This represents the pixel value after lightweight smoothing.

4. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 3, characterized in that: Step 1 also includes S1.6: During the training phase, the device category labels and bounding box coordinates in the annotation file are synchronously transformed to correspond to the preprocessed network input image; specifically as follows: During the inference phase, only the image scaling and padding parameters are retained for coordinate inverse mapping in step 5; Let the first part of the original infrared image be... The category label for each target is Its actual labeled bounding box is represented as: （26）； In equation (26), and They represent the first The pixel coordinates of the top left and bottom right corners of the target bounding box in the original infrared image; Indicates the first The true bounding box of each target; Based on the scaling ratio obtained in step 1.3 and fill offset , The coordinates of the bounding box after synchronous transformation are: （27）； （28）； （29）； （30）； In the above formula, , These represent the coordinates of the top left and bottom right corners of the bounding box in the network input image after synchronous transformation; Accordingly, the bounding box after synchronous transformation is represented as: （31）； This represents the bounding box after synchronous transformation; The equipment category label remains unchanged during geometric transformation: （32）； This represents the category label after the geometric transformation, and the category label remains unchanged before and after the transformation. When the training annotations use normalized center point coordinates and width / height, further calculations can be performed: （33）； （34）； （35）； （36）； In the formula, They represent the first The normalized coordinates of the center point of each target bounding box relative to the width and height of the network input image; and These represent the normalized bounding box width and height, respectively; and These represent the width and height of the input image to the network, respectively.

5. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 4, characterized in that: Step 2, which involves constructing a lightweight backbone feature extraction network, specifically includes: S2.1: Based on the single-stage target detection framework, a lightweight backbone feature extraction network for infrared images is constructed. The lightweight backbone feature extraction network consists of an initial convolutional layer and multiple levels of lightweight inverse residual feature units, which are used to extract multi-scale features from the input infrared image, from shallow edge texture to deep semantic information. S2.2: Input the image to be identified obtained in step 1 into the initial convolutional layer to obtain the initial infrared feature map; the initial convolutional layer is used to complete the mapping of the infrared image from pixel space to feature space, and to initially extract the thermal edge, local texture and regional grayscale difference information of the power equipment; specifically as follows: Let the network input infrared image obtained in step 1 be: （37）； In the formula, This represents the infrared image tensor input to the network. Indicates the number of input channels; , These represent the height and width of the network input image, respectively. The initial infrared feature map is obtained through the initial convolutional layer. , is represented as: （38）； In the formula, This represents the infrared feature map output by the initial convolutional layer; Represents a non-linear activation function; This indicates a batch normalization operation; Indicates the initial convolutional layer kernel parameters; * represents the convolution operator. Indicates the bias parameter; No. The output channels are located in space. The eigenvalues ​​at that location are represented as: （39）； In the formula, Represents the initial feature map. The output channels are located in space. Eigenvalues ​​at; Indicates the output channel number; Indicates the input channel number; Indicates the initial convolution kernel size; , Indicates the position index inside the convolution kernel; Indicates the connection of the first Input channel and the first Output channel, convolution kernel The weighting parameters of the location; The first input tensor represents the... One channel; Indicates the convolution stride; Indicates the fill size; Indicates the first The bias corresponding to each output channel; S2.3: The initial infrared feature map is input into a multi-level lightweight inverse residual feature unit to obtain infrared feature maps with different spatial resolutions. The shallow high-resolution feature map is used to preserve device edges and small target textures, the mid-level semantic feature map is used to express the device's structural morphology, and the deep low-resolution feature map is used to provide semantic information about the device category. Specifically: Let the initial convolutional layer output be , No. The lightweight feature extraction unit is denoted as The multi-level feature extraction process can then be represented as: （40）； In equation (40), Indicates the number of levels in a lightweight feature extraction network; and They represent the first The input and output feature maps of the level feature extraction unit; Indicates the first Lightweight feature transformation; This represents the output of the initial convolutional layer, i.e., the 0th level feature; Indicates a network level index; Three scale features were selected from different network depths and denoted as follows: （41）； （42）； （43）； In the formula, , and These represent infrared feature maps at three scales with step sizes of 8, 16, and 32, respectively. , and These represent the number of channels in the corresponding scale feature map; and This indicates the height and width of the input image to the network; When the network input size is At that time, the corresponding spatial dimensions are as follows: （44）； Its multi-scale output modes of P3 / 8, P4 / 16, and P5 / 32 correspond to the three detection scale settings of the YOLOv11 detection structure; S2.4: The lightweight inverse residual feature unit includes a channel expansion layer, a spatial modeling layer, a channel projection layer, and a residual connection layer. The channel expansion layer uses 1×1 convolutions to expand the input features, enhancing the feature representation dimension. The spatial modeling layer uses depthwise convolutions or depthwise separable convolutions to extract local infrared spatial features. The channel projection layer uses 1×1 convolutions to compress the number of channels, reducing computational cost. When the input and output feature sizes and number of channels are the same, the residual connection layer adds the input features and the projected features to enhance gradient propagation and reduce information loss caused by lightweight compression. Specifically: Let the input features of the lightweight reciprocal residual feature unit be: （45）； In the formula, This represents the input features of the lightweight inverse residual feature unit; This represents the number of channels for the input feature; , These represent the height and width of the feature map, respectively. First, optional channel-wise spatial feature extraction is performed: （46）； In the formula, This indicates channel-wise convolution, where each channel undergoes spatial convolution independently. , These represent the kernel size and stride of the preceding channel-wise convolution, respectively. Then adopt Channel expansion via pointwise convolution: （47）； In the formula, This indicates the characteristics after channel expansion; Indicates the use of channel expansion Pointwise convolution; The expanded number of channels is expressed as: （48）； In equation (48), Indicates the channel spread factor; This represents the number of feature channels after expansion; Indicates the rounding operation; Then perform channel-wise spatial convolution: （49）； In the formula, This represents the intermediate features obtained by channel-wise spatial convolution after channel expansion; , These represent the kernel size and stride of the channel-wise spatial convolution after channel expansion, respectively. Then through Pointwise convolution completes channel projection: （50）； In the formula, Indicates Features obtained after pointwise convolution channel projection (compression); This represents pointwise convolution used for channel projection; When the input and output feature spaces have the same size and the same number of channels, residual connection is used: （51）； In the formula, This represents the final output feature of the lightweight inverse residual feature unit; otherwise: （52）； When the input and output dimensions or number of channels are inconsistent, residual connections are not enabled, and the projected features are output directly. S2.5: The lightweight backbone feature extraction network adopts a MobileNetV4-type backbone structure and selects a convolutional small model configuration as the backbone parameter configuration; by decomposing standard convolution operations through channel-wise spatial convolution and point-wise channel-wise convolution, infrared spatial texture extraction and cross-channel information fusion are performed separately; specifically as follows: In standard convolution, the first The eigenvalues ​​of each output channel are represented as follows: （53）； In equation (53), Represents the first... One channel; The first feature map is the one whose output feature map is represented by the first feature map. One channel; Indicates the standard convolution kernel parameters; Indicates the number of input channels; Indicates the side length of the convolution kernel; When using depthwise separable convolution, spatial convolution is first performed on each input channel separately: （54）； In the formula, Indicates the first Intermediate features obtained by channel-wise spatial convolution of each input channel; Indicates the first The depthwise convolution kernel parameters corresponding to each channel; Then adopt Pointwise convolution completes cross-channel information fusion: （55）； In the formula, Indicates by the first The intermediate channel is mapped to the first... Each output channel Pointwise convolution parameters; Therefore, the standard convolution parameter count is: （56）； In the formula, The number of parameters representing a standard convolution; Indicates the number of channels in the output feature map; This represents the number of spatial parameters of a single convolutional kernel; The number of parameters in a depthwise separable convolution is: （57）； In the formula, The number of parameters representing a depthwise separable convolution; the first term The second term represents the number of parameters in the channel-wise spatial convolution; express Pointwise convolution parameter count; S2.6: Let the number of input feature map channels be... The number of channels in the output feature map is The kernel size is The feature map space size is Standard convolution computational cost Depthwise separable convolution The computational load is expressed as follows: （58）； （59）； In the formula, and These represent the height and width of the output feature map, respectively; and These represent the number of channels in the input feature map and the output feature map, respectively. Indicates the side length of the convolution kernel; Indicates the computational cost of standard convolution; Equation (58) represents the computational cost of depthwise separable convolution; Equation (59) represents the computational cost of standard convolution; where, in Equation (59), the first term represents the computational cost of channel-wise spatial convolution, and the second term represents... The computational cost of cross-channel information fusion via pointwise convolution; S2.7: To ensure that the output features of the backbone network can be adapted to subsequent detection branches, the infrared feature maps of different scales output by the multi-level lightweight inverse residual feature units are numbered according to scale, and are designated as P3, P4 and P5. Among them, P3 retains higher resolution and more detailed information, and is used for small-scale substation target recognition; P4 takes into account both spatial information and semantic expression; P5 contains stronger semantic information and is used for larger-scale equipment targets and overall structure recognition.

6. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 5, characterized in that: Step 3 specifically includes: S3.1: Input the infrared feature map of at least one scale obtained in step 2 into the channel response recalibration module; the channel response recalibration module is set between the output end of the lightweight backbone feature extraction network and the input end of the detection branch, adaptively evaluates the importance of different channels in the infrared feature map, and enhances the target-related feature response of the device without significantly increasing the number of parameters and computation. S3.2: For any scale of input infrared feature map ,in: Indicates the number of channels. and These represent the height and width of the feature map, respectively. The channel response recalibration module processes the data using a parallel approach of preserving the original feature bypass and generating channel weight branches; the input feature map... On the one hand, it is directly passed along the bypass branch to preserve the original spatial structure information; on the other hand, it is input into the global average pooling layer (GAP) to extract the global response statistics of each channel. For the Each channel undergoes global average pooling, and its channel descriptor value Represented as: （60）； In equation (60), Indicates the input infrared feature map The Middle The spatial location of each channel eigenvalues ​​at This represents the spatial position index along the height direction of the feature map, with a value range of [value range missing]. ; This represents the spatial position index along the width direction of the feature map, with a value range of [value range missing]. ; Indicates the first The channel description values ​​obtained after global average pooling of each channel; This represents the total number of spatial locations in the feature map; S3.3: Obtaining the channel description vector Then, based on the number of channels in the input feature map Adaptive determination of one-dimensional convolution kernel size , The convolution kernel size With the number of channels Correspondingly, infrared feature maps of different channel sizes adopt matching local channel interaction ranges; the adaptive convolution kernel size is expressed as: （61）； In equation (61), and Preset adjustment parameters; This indicates taking the nearest odd number; Indicates the size of the one-dimensional convolution kernel; Indicated by the number of channels The adaptive convolution kernel size mapping function is the independent variable; This indicates the number of channels in the input feature map; Represents the logarithmic function with base 2; S3.4: The one-dimensional convolution kernel size determined using S3.3 , channel description vector Perform one-dimensional convolution operations to establish local dependencies between adjacent channels and obtain channel interaction response vectors. The calculation process is expressed as follows: （62）； In equation (62), Indicates the kernel size as One-dimensional convolution operation; Represents the channel interaction response vector; This represents the channel description vector composed of all channel description values. ; S3.5: The channel interaction response vector obtained in S3.4 Inputting the Sigmoid activation function yields the channel weight vector. The calculation process is expressed as follows: （63）； In equation (63), This represents the Sigmoid activation function; This represents the generated channel weight vector; S3.6: The channel weight vector obtained in S3.5 Compared with the original infrared feature map preserved by bypass in S3.2 Perform channel-by-channel multiplication to obtain the recalibrated effective infrared feature map. Its calculation process is expressed as equation (64): （64）； In the formula, Indicates the first The channel weights corresponding to each channel; Indicates the first Output characteristics after recalibration of each channel; The original feature map representing bypass preservation. One channel; Or expressed as equation (65): （65）； In the formula, Indicates the first Output characteristics after recalibration of each channel; This indicates a channel-by-channel multiplication operation; S3.7: For infrared feature maps of different scales, execute S3.2~S3.6 respectively to independently generate channel weights for the corresponding scales and complete feature recalibration; when the input feature map is a shallow high-resolution feature, the channel response recalibration module focuses on enhancing device edges, small target textures, and local thermal structure information; when the input feature map is a mid-level feature, it focuses on enhancing device morphology and local structural information; when the input feature map is a deep low-resolution feature, it focuses on enhancing semantically related channels for device categories. S3.8: Output the recalibrated infrared effective feature maps at each scale to the subsequent detection branch.

7. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 6, characterized in that: Step 4 specifically includes: S4.1: Input the multiple scale infrared effective feature maps obtained in S3 into the single-stage detection branch, which specifically includes a classification prediction sub-branch, a target confidence prediction sub-branch, and a bounding box regression sub-branch; wherein, the classification prediction sub-branch is used to determine the equipment category to which the candidate target belongs, the target confidence prediction sub-branch is used to determine whether there is a substation equipment target in the candidate region, and the bounding box regression sub-branch is used to predict the position of the equipment target in the image; The classification prediction sub-branch outputs the probability value of each candidate target belonging to categories such as surge arresters, high-voltage bushings, voltage transformers, line hardware, and insulators; the target confidence prediction sub-branch outputs the probability that the candidate region contains substation equipment targets; the bounding box regression sub-branch outputs the coordinates of the bounding box center point, width, and height; as detailed below: For the Each scale feature map The output of the detection branch is represented as: （66）； In equation (66), This indicates three detection scales; Indicates the first Each scale corresponds to a detection head; This represents the bounding box regression output; This represents the target confidence prediction output; This represents the category prediction output; For feature map location First The probability of a device class is expressed as follows: （67）； in, （68）； In the formula, Indicates the output of the detection head. Class prediction value; Indicates the probability of the corresponding category; ,in Indicates the total number of equipment categories; The predicted device category is: （69）； In the formula, Indicates the location of the feature map Predicted device category at the location; Represents the probability of taking the class. Largest category number ; The corresponding target confidence level is defined as: （70）； In the formula, Indicates the location of the feature map The predicted confidence level of the target at the location; This represents the Sigmoid activation function; This indicates the confidence level that the corresponding candidate region contains the target of the power equipment. Suppose that the bounding box regression yields the predicted bounding box: （71）； This can then be further converted into a center point and width / height format: （72）； （73）； （74）； （75）； In the formula, and These represent the coordinates of the top-left and bottom-right corners of the predicted bounding box, respectively. Indicates the coordinates of the center point of the bounding box; and These represent the width and height of the bounding box, respectively. Therefore, the first The candidate identification results are uniformly represented as follows: （76）； In equation (76), , Indicates the first One candidate identification result; Indicates the predicted device category; Indicates the confidence level of the target; Indicates the coordinates of the predicted bounding box center point; and These represent the width and height of the predicted bounding box, respectively. For each predicted target, the detection branch outputs its class probability, target confidence, and bounding box coordinates. Each scale feature map corresponds to three types of outputs: classification prediction, confidence prediction, and bounding box regression. S4.2: During the training phase, the lightweight backbone feature extraction network, channel response recalibration module, and detection branch are jointly trained based on the category labels and bounding box labels in the infrared image dataset of substation equipment. Training losses include classification loss, object confidence loss, and bounding box regression loss: The detection branch uses a joint loss function to jointly optimize classification prediction, target confidence prediction, and bounding box regression. The classification loss is used to constrain the consistency between the predicted category and the true category, the target confidence loss is used to determine whether the candidate region contains the substation target, and the bounding box regression loss is used to constrain the degree of overlap between the predicted box and the true box. Let the predicted bounding box be The actual labeled bounding box is The intersection and union ratio of the two Represented as: （77）； In equation (77), This represents the intersection area of ​​the predicted bounding box and the ground truth bounding box. The intersection-union ratio represents the area of ​​the union of the two objects; based on the intersection-union ratio, the bounding box regression loss is... As shown in equation (78): （78）； Total loss function for detection branches As shown in equation (79): （79）； In equation (79), Represents classification loss, Indicates the target confidence loss. This represents the bounding box regression loss; , , These represent the weight coefficients of each loss term; through the joint loss function described above, the detection branch can improve the matching degree between the predicted bounding box and the real target region while completing device category identification. S4.3: During the inference phase, the detection branch generates candidate recognition boxes based on the infrared effective feature map output in step 3, and obtains the device category, confidence level and bounding box coordinates of each candidate target.

8. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 7, characterized in that: Step 5 specifically includes: S5.1: Read the candidate recognition results obtained in 4.3, wherein the candidate recognition results include the device category, target confidence level, and bounding box coordinates of the candidate target; establish a candidate box set for all candidate recognition results in the same infrared image, denoted as: （80）； In equation (80), Represents the set of candidate recognition results. Indicates the first One candidate identification result, Indicates the number of candidate recognition results; S5.2: Confidence screening of candidate identification results: Let the first The target confidence level for each candidate identification result is: The preset reliability threshold is ,when When, delete the candidate recognition result; when If necessary, retain the candidate identification result for subsequent processing; S5.3: The candidate identification results after confidence screening are grouped according to equipment category; for different categories such as surge arresters, high-voltage bushings, voltage transformers, line fittings, and insulators, candidate box sorting and duplicate box suppression are performed separately, as follows: Let the set of candidate identification results after confidence level filtering be: （81）； In the formula, This represents the set of candidate identification results after confidence level filtering; This indicates the total number of candidate boxes retained after filtering; Among them, the The candidate identification results are represented as follows: （82）； In the formula, Indicates the first Predicted bounding boxes of candidate targets; Indicates the predicted device category; Indicates the confidence level of the target; This indicates the total number of candidate boxes retained after confidence level filtering; For the For each device class, create a corresponding set of candidate frames: （83）； In the formula, This indicates the total number of categories of power equipment; Indicates the first The set of candidate boxes corresponding to the device class; For sets Candidate results in the data are ranked according to confidence level. Arranged from highest to lowest; let the sorted index sequence be: （84）； In the formula, Indicates the first A sorted index sequence of candidate boxes arranged in descending order of confidence. This indicates an operation that returns the index of the sorted element; The negative sign indicates that the order is descending (from highest to lowest confidence level). The sorted candidate box set is then represented as: （85）； In the formula, This represents the sorted set of candidate boxes; Represents the original index of the 1st, 2nd...th position in the sorted sequence; Indicates the first The number of candidate boxes corresponding to the device class; And satisfy: （86）； In the formula, Indicates the first The number of candidate boxes corresponding to the device class; This indicates an operation that returns the index of the sorted element; S5.4: Perform non-maximum suppression processing on candidate boxes within the same device category; specifically: First, the candidate boxes are sorted from highest to lowest target confidence level, and the candidate box with the highest confidence level is selected as the current retained box; the details are as follows: Let the first The current set of candidate boxes to be processed is ,in The first nonmaximum suppression represents the... In the next iteration, the index of the candidate box with the highest confidence is represented as: （87）； The corresponding candidate results Add to the final reserved set : （88）； In the formula, Indicates the first The index of the candidate box with the highest target confidence in the next iteration; Indicates the first The target confidence level of each candidate box; Indicates the first The first iteration has already retained the... A collection of candidate boxes; This represents the set union operation (combining the current best box into sets). (Add to the reserved set) Represents the updated retainable set; Then, the intersection-union ratio (IUR) between the remaining candidate boxes and the currently retained box is calculated. If the IUR is greater than a preset threshold, it is considered that the two correspond to the same substation equipment target, and the candidate boxes with lower confidence are judged as duplicate candidate boxes and deleted; specifically as follows: Let the currently retained box be: （89）； In the formula, This indicates the currently reserved box (the box with the highest confidence level); , Indicates the coordinates of its top left and bottom right corners; The candidate boxes to be compared are: （90）； In the formula, Indicates the candidate boxes to be compared; , Indicates the coordinates of its top left and bottom right corners; The width and height of the intersection region are represented as follows: （91）； In the formula, This indicates the width of the intersection area between the two frames; : The right boundary of the intersection (take the smaller of the right boundaries of the two boxes); The left boundary of the intersection is represented by the larger of the left boundaries of the two boxes; outer layer This indicates that the intersection width is 0 when the two boxes do not intersect. （92）； In the formula, Indicates the height of the intersection area of ​​the two frames; The area of ​​intersection is: （93）； In the formula, This represents the area of ​​the intersection of the two frames; The areas of the two candidate boxes are as follows: （94）； In the formula, This indicates the area of ​​the currently reserved frame; （95）； In the formula, This represents the area of ​​the candidate boxes to be compared; The intersection, union, and ratio of the two can be expressed as: （96）； In the formula, This indicates the intersection-union ratio (IoU) of the reserved box and the box to be compared. Let the nonmaximum suppression threshold be... ,when: （97）； and: （98）； In the formula, Indicates the confidence level of the current reserved box; Indicates the confidence level of the candidate boxes to be compared; Then the candidate box Determined as the currently retained frame The corresponding duplicate prediction boxes are removed from the set to be processed; The candidate box set after one iteration is represented as: （99）； In the formula, This represents the set of candidate boxes remaining after one iteration. To perform a set difference operation, first remove the currently retained box. ;condition This means that only candidate boxes whose intersection-union ratio with the currently retained boxes does not exceed the threshold are retained, and duplicate boxes are deleted; Repeat the above process until: （100）； In the formula, This indicates an empty set, meaning the iteration terminates when the set to be processed is empty; at this time, That is, the first The set of candidate recognition results retained by the device after nonmaximum suppression; Repeat the above process until all candidate boxes in this category have been processed.

9. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 8, characterized in that: Step 5 further includes S5.5: performing coordinate inverse mapping on the bounding box retained after nonmaximum suppression: Based on the image scaling ratio, fill size, and original image size recorded in step 1, the bounding box coordinates under the network input size are converted back to the original infrared image coordinates to obtain the actual position of the device target in the original image; specifically as follows: Let the coordinates of the predicted bounding box retained after nonmaximum suppression in the network input image be: （101）； In the formula, This represents the coordinates of the predicted bounding box preserved by NMS in the coordinate system of the network input image; , This represents the coordinates of the top-left and bottom-right corners of the predicted bounding box in the network input image; Based on the image scaling ratio recorded in step 1.3 And the fill offset in the horizontal and vertical directions. and The predicted bounding box is then mapped inversely to the original infrared image coordinate system, as follows: （102）； In the formula, This represents the predicted bounding box in the coordinate system of the network input image; This represents the bounding box after inversion onto the original infrared image; and These represent the coordinates of the top-left and bottom-right corners of the predicted bounding box in the network input image, respectively. This represents the scaling factor of the image recorded in step 1.3; and These represent the fill offsets of the image in the horizontal and vertical directions, respectively. S5.6: Output the final infrared image recognition results of the substation equipment: The recognition results include at least the device category, bounding box coordinates, target confidence score, and image number; Based on actual inspection needs, it also includes data collection time, inspection points, equipment number, data collection terminal number, and identification result storage path; S5.7: Associate the final recognition result with the original infrared image; for scenarios requiring visualization, overlay the device category name, target confidence score, and bounding box position onto the original infrared image to form a visual recognition map; details are as follows: For scenarios requiring data retrieval, the recognition results can be output to the inspection management platform in the form of structured data; Let the first The final identification result is: （103）； In the formula, Indicates the first The final identification result is a triplet: bounding box, category, and confidence score. in: （104）； Indicates the final bounding box; Indicates the equipment category; Indicates the confidence level of the target; Definition of the first The set of boundary pixels of each bounding box is: （105）； In the formula, Indicates the first The set of boundary pixels of a bounding box; The first set consists of two horizontal edges, one above and one below. Pick or , It takes values ​​continuously between the left and right boundaries; The second set consists of two perpendicular edges, left and right. Represents the union; For all final bounding boxes, define a visual boundary mask: （106）； In the formula, Represents the visual boundary mask; This represents the union of all final bounding box pixels; The image after the bounding boxes are superimposed is represented as: （107）； In equation (107), Represents the original infrared image; This represents a visualized image after the recognition results have been overlaid. Represents the bounding box mask; Indicates the relationship with the first Preset display values ​​or display colors corresponding to this type of device; No. The visual text labels corresponding to each target are represented as follows: （108）； In equation (108), This represents the visual text label corresponding to the i-th target; Indicates based on category number The device category name obtained from the query; This indicates the target confidence level after formatting according to a preset number of decimal places; This represents a string concatenation operation; Text Label It can be set near the top left corner of the bounding box: （109）； In the formula, Represents text label The display position coordinates are set at the top left corner of the bounding box. nearby; This results in a visual recognition outcome that includes the device category name, target confidence level, and spatial location. S5.8: The data fields of the final recognition result are set according to a table; the table includes field names, field meanings and data types, which are used to unify the storage format and interface output format of the recognition results.

10. The lightweight infrared image recognition method for substation equipment oriented towards end-side inspection according to claim 9, characterized in that: It also includes step 6: constructing a lightweight infrared image recognition system for substation equipment, deploying the recognition processes corresponding to steps 1 to 5 on the end-side inspection platform to realize infrared image acquisition, preprocessing, lightweight recognition, result output, and inspection data management; specifically including: S6.1: An infrared image input interface is set in the end-side inspection platform. The infrared image input interface is connected to the UAV infrared camera, the inspection robot infrared imaging equipment, the fixed infrared monitoring terminal or the handheld infrared imaging equipment to receive infrared images collected during the substation inspection process and send the infrared images into the subsequent image processing flow. S6.2: Establish a preprocessing process for the input infrared image. The preprocessing process includes image size normalization, temperature grayscale mapping, noise suppression, pixel normalization, and coordinate mapping parameter recording. After preprocessing, the original infrared image is converted into an image to be recognized that meets the input requirements of the lightweight recognition network, while retaining the image scaling ratio and padding information so that the subsequent recognition results can be mapped back to the original image coordinates. S6.3: The image to be identified is sent to the lightweight feature extraction unit, which performs the backbone feature extraction process in step 2 to obtain infrared feature maps with different spatial resolutions. This unit reduces the number of parameters and computations in the end-side inference process through the lightweight backbone network, providing a multi-scale feature basis for subsequent target recognition of the device. S6.4: After the lightweight feature extraction unit, a channel response recalibration unit is set up to perform the channel weight calculation and feature recalibration processing in step 3 on the multi-scale infrared feature map; through this processing, the channels related to the device thermal edge, main body contour and structural texture are enhanced, while the channel responses corresponding to the background hot spot and weakly correlated areas are suppressed, thereby obtaining an effective infrared feature map; S6.5: Input the effective infrared feature map into the recognition output unit, which then performs candidate box generation, confidence level filtering, non-maximum suppression, and coordinate inverse mapping processing, and outputs the infrared image recognition result of the substation equipment; the recognition result includes equipment category, equipment location, and target confidence level. S6.6: Store the equipment category, bounding box coordinates, target confidence score, inspection task number, acquisition time, inspection point, equipment number, and original infrared image; for scenarios requiring manual review, the recognition box, category name, and confidence score can be overlaid on the original infrared image to generate a visual recognition result; for scenarios requiring platform access, the recognition result can be converted into structured data and output to the inspection management platform. S6.7: During the operation of the end-side inspection platform, the infrared image input interface receives inspection images in real time. The preprocessing process generates images to be identified. The lightweight feature extraction unit and the channel response recalibration unit sequentially complete feature extraction and effective feature enhancement. The identification output unit generates equipment identification results. The inspection data association mechanism completes result storage and display. For continuous frame infrared images, the system executes the above process frame by frame according to the acquisition order and outputs the substation equipment identification results in each frame of the image. S6.8: Configure the input image size, model inference accuracy, confidence threshold, non-maximum suppression threshold, and result output fields according to the computing power, storage conditions, and real-time requirements of the edge inspection platform.

Images