A composite insulator detection method based on multi-scale infrared enhancement

Abstract

The application provides the technical field of infrared insulator detection, and provides a composite insulator detection method based on multi-scale infrared enhancement, takes YOLOV11 as a basic framework, embeds a C3k2_CG module in a YOLO main stem shallow layer and a middle layer, adopts a SPPF_SE module in a spatial pyramid pooling stage of the YOLO main stem high layer, embeds an ASF_YOLO feature fusion unit between the YOLO main stem and a neck, inserts a SimAM parameter-free attention module on each scale feature map of a YOLO detection head, and enhances multi-scale features; in the pre-training stage of the infrared composite insulator detection model, an infrared image special enhancement module IRAE is introduced. The application significantly improves the distinguishability and boundary definition of the target and the background in a complex environment such as low contrast, thermal noise and temperature change, enhances the multi-scale modeling and positioning capability, reduces the false detection and the missed detection, and improves the precision, the recall rate and the overall reliability.

Description

Technical Field

[0001] This application belongs to the field of infrared insulator detection technology, specifically relating to a composite insulator detection method based on multi-scale infrared enhancement. Background Technology

[0002] Composite insulators are key equipment for ensuring the safe and stable operation of power grids. Due to their long-term exposure to complex environments such as high voltage, strong ultraviolet radiation, salt spray, acid rain, wind and sand, high humidity, and temperature fluctuations, while also enduring mechanical stress, vibration, and aging, their operating status directly affects the insulation level of lines and the safe and stable operation of the power grid. Monitoring their condition is crucial for preventing power accidents. Traditional manual inspections rely on the experience of maintenance personnel to identify visual abnormalities. While low-cost, this method is significantly affected by lighting and viewing angle, making it difficult to detect early, subtle defects. Ultraviolet imaging, partial discharge, and ultrasonic testing can identify and locate corona and partial discharge. Online monitoring based on leakage current and dielectric loss provides continuous and quantifiable operating indicators. Currently, infrared thermal imaging is the most widely used method due to its non-contact and high efficiency, often used in drone or vehicle-mounted inspections, and is highly sensitive to abnormal temperature rises.

[0003] However, the detection of composite insulator targets in infrared images still faces many challenges. Composite insulators are often slender and far apart, with few individual target pixels; the shading of the skirt structure, the reflection of the hardware, and the large changes in posture make it difficult to detect small targets, resulting in irregular boundaries and unclear outlines; infrared images generally have low contrast, high noise, blurred edges, uneven background temperature distribution, and frequent reflections and glare, making it difficult to separate the target from the background, and the features are unstable, leading to false positives or false negatives. All of these directly affect the accuracy, stability, and usability of subsequent defect detection. Summary of the Invention

[0004] This application provides a composite insulator detection method based on multi-scale infrared enhancement to solve or partially solve the problems mentioned in the background art.

[0005] This application provides a method for detecting composite insulators based on multi-scale infrared enhancement, including: Based on the YOLOv11 framework, an infrared composite insulator detection model was constructed to detect composite insulators in infrared images. The specific construction of the infrared composite insulator detection model is as follows: The C3k2_CG module is embedded in the shallow and middle layers of the YOLO backbone to enhance the joint modeling of local textures and surrounding context; In the spatial pyramid pooling stage of the YOLO backbone, the SPPF_SE module is used to recalibrate the channels of the high-level features and perform multi-scale pooling and stitching. An ASF_YOLO feature fusion unit is embedded between the YOLO backbone and the Neck to extract and fuse multi-scale features. SimAM nonparametric attention modules are inserted into the feature maps at each scale of the YOLO detection head to enhance the multi-scale features; In the pre-training stage of the infrared composite insulator detection model, an infrared image enhancement module (IRAE) is introduced to enhance the infrared contrast and structural details of the samples.

[0006] Preferably, the infrared image enhancement module (IRAE) includes a contrast adaptive enhancement module, a thermal noise simulation module, a temperature change simulation module, and a boundary preservation filtering module. The functions of each module are as follows: The contrast adaptive enhancement module employs an adaptive strategy to dynamically adjust enhancement parameters based on the average brightness value of the image in order to achieve adaptive contrast enhancement of infrared images. The thermal noise simulation module uses a composite noise model, which superimposes Gaussian noise onto the original infrared image. Gaussian thermal noise simulates random noise generated by thermal motion. The temperature change simulation module simulates the imaging effect under different temperature conditions; The boundary preservation filtering module employs bilateral filtering and unsharpening masking techniques to maintain the clarity of the target boundary while performing blurring or sharpening processing.

[0007] Preferably, the contrast adaptive enhancement module processes the input image using a mathematical model. If the image is a dark image with a mean <100, the stretching coefficient α associated with contrast enhancement is [1.3, 1.8], and the stretching coefficient β associated with brightness enhancement is [10, 30]. If the image is a bright image with a mean ≥100, the stretching coefficient α associated with contrast enhancement is [1.1, 1.5], and the stretching coefficient β associated with brightness enhancement is [0, 15]. The mathematical model is as follows: in, I in (x,y) For input pixels, and For local mean and standard deviation, α、β These are the stretching parameters; The thermal noise simulation module employs a composite noise model in the original infrared image. I in (x,y) Gaussian noise is added on top of the existing structure, as shown in the following formula: in, With a mean of 0 and a variance of Gaussian noise; The temperature change simulation module simulates the effect of temperature increase or decrease by linearly adjusting the overall pixel values ​​of the input image, as shown in the following formula: in, ΔT The change in temperature k This is the coefficient representing the effect of temperature changes on pixels; The boundary preservation filter module, through and The weighted average preserves the sharpness of the target edges during denoising or sharpening, as shown in the following formula: in, I in (x,y) For input pixels, For a spatial Gaussian kernel, The Gaussian kernel for pixel value differences, Ω is the normalization coefficient, and Ω is the neighborhood window.

[0008] Preferably, the C3k2_CG module replaces the traditional convolutional bottleneck in the bottleneck branch of the original C2f structure with ContextGuidedBlock. ContextGuidedBlock concatenates local features with contextual features, and then introduces global channel attention for joint modeling. The specific method is as follows: After dimensionality reduction of the input features using 1×1 convolution, they are fed into the local feature extraction branch using standard 3×3 channel convolution and the surrounding context extraction branch using dilated channel convolution, respectively. The two outputs are spliced ​​together along the channel dimension and then combined using batch normalization and PReLU nonlinear transformation to form a joint feature. A global context extraction unit consisting of global average pooling and fully connected layers is used to generate channel attention weights for joint features, and adaptive recalibration is performed on each channel.

[0009] Preferably, the SPPF_SE module utilizes its SELayerV2 submodule to perform channel recalibration on high-level features and execute multi-scale pooling and concatenation, as follows: Global adaptive pooling and channel weighting are applied to the input high-level features; The input feature map is first subjected to global average pooling to compress the spatial dimension into channel description vectors. Then, the channel information is nonlinearly transformed through a multi-branch fully connected layer and spliced ​​and aggregated in the channel dimension. An excitation operation is performed on the aggregated channel description vectors. Normalized channel attention weights are generated through a fully connected layer and a sigmoid activation, and then multiplied element-wise with the input feature map to complete channel recalibration. The recalibrated features are then fed into the residual convolution branch F(·), and finally added to the input feature x using residuals. The calculation process is expressed by the following formula: Where x represents the input feature map, Sq This represents the description of the multi-branch path obtained from the squeeze operation. ∑Sq This represents the aggregation of descriptions of various branch passages. Ex This represents the channel attention weights obtained from the excitation mapping. F This indicates that the input features are weighted and labeled channel by channel.

[0010] Preferably, the specific method for extracting multi-scale fusion features by the ASF_YOLO feature fusion unit is as follows: The SSFF submodule uses Zoom_cat and Concat operations to perform scale alignment and multi-scale fusion of features in layers P3, P4 and P5. The TFE submodule utilizes the ScalSeq operation to decompose and enhance the fused features at three scales: large, medium, and small. The CPAM submodule introduces cross-scale attention at the Add position to integrate information from SSFF and TFE. The SimAM parameterless attention module embeds the SimAM parameterless attention weight allocation mechanism for feature enhancement at three different scales: P3, P4, and P5. Finally, it outputs multi-scale detection results through three Detect modules.

[0011] Preferably, the SSFF module effectively fuses feature maps from layers P3, P4, and P5 to capture different spatial information, covering target features of various sizes and shapes. It also normalizes the feature maps from different layers to achieve a uniform spatial resolution. Finally, the processed multi-scale feature maps are stacked as input for subsequent convolution operations. The feature maps from different layers are convolved with a series of Gaussian kernels with increasing standard deviations, as expressed by the formula: Where f(w,h) represents a two-dimensional input image with width w and height h, respectively. It consists of a series of convolutions using a two-bit Gaussian filter. The smoothing process generates σ, which is the standard deviation scaling parameter of the two-dimensional Gaussian filter used for convolution.

[0012] Preferably, the TFE module processes feature maps of different sizes by adjusting the number of feature channels, using a hybrid pooling structure for downsampling, using transposed convolution for upsampling, and employing an attention weight allocation mechanism. It then fuses feature maps of different sizes and concatenates them along the channel dimension, as shown in the following formula: in, The feature map output by the TFE module. , , These represent feature maps of large, medium, and small sizes, respectively. resolution and The same, and the number of channels is 3 times.

[0013] Preferably, the CPAM module consists of a channel attention network Input1 that receives TFE input, an input position attention network Input2 that receives channel attention networks and SSFF outputs, and introduces an adaptive attention weight allocation mechanism based on 1D convolution. Channel attention is achieved by considering the local cross-channel interactions of each channel and its K nearest neighbors, and an adaptive relationship between the channel dimension C and the convolution kernel K is established, as follows: γ Set b to 2, and b to 1. γ and b This is a scaling parameter used to control the ratio of the convolution kernel size k to the channel dimension C. It is an odd number of nearest neighbors.

[0014] Preferably, the SSFF module introduces a synergistic effect of positional attention weight allocation mechanism and channel attention weight allocation mechanism. Based on average pooling (pw) and max pooling (ph), feature distribution analysis and weight allocation are performed to preserve the spatial structure information of the feature map, thereby achieving effective aggregation of spatial encoding information. Its mathematical expression is as follows: Where E(w,j) and E(i,h) are the input feature maps at position (i,j); When generating position attention coordinates, the CPAM module applies a concatenation operation to the horizontal and vertical axes to obtain the position attention coordinates, and then uses a split operation to pair the position-dependent feature maps, as shown in the following formula: Wherein, P( , ) is the output of the position attention coordinates. and These are the width and height of the segmented output, respectively; Finally, the output of the CPAM module is defined by the following formula: Here, E represents the weights of channel attention and position attention.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: The composite insulator detection method based on multi-scale infrared enhancement proposed in this application introduces C3k2_CG and SPPF_SE modules on the backbone side, and adds SimAM parameterless attention to the ASF_YOLO multi-scale feature fusion unit and the detection head on the neck side. This achieves collaborative optimization from shallow local texture modeling, high-level global context modeling to cross-scale feature fusion and attention allocation. Combined with IRAE infrared enhancement in the preprocessing stage, it significantly improves the distinguishability and boundary clarity of the target and background in complex environments such as low contrast, thermal noise and drastic temperature changes, enhances multi-scale modeling and localization capabilities, reduces false detections and false negatives, and improves precision, recall and overall reliability.

[0016] Meanwhile, this method has stronger cross-scenario adaptability and robustness through the synergistic effect of multiple modules. It achieves end-to-end optimization from preprocessing, backbone feature extraction, spatial pyramid pooling, multi-scale feature fusion to attention allocation. It can be seamlessly integrated with the existing power inspection process, reduce the workload of manual re-inspection, improve inspection efficiency, and provide high-quality detection results for subsequent defect diagnosis and risk assessment, thereby strongly supporting the safe and stable operation of the power system. Attached Figure Description

[0017] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is a schematic diagram of the C3k2_CG model structure of this application. Figure 2 This is a schematic diagram of the ContextGuidedBlock model structure in this application. Figure 3 This is a schematic diagram of the SPPF_SE model structure in this application. Figure 4 This is a schematic diagram of the ASF_YOLO model structure in this application. Figure 5This is a schematic diagram of the three-dimensional weighted attention weight allocation mechanism of this application. Figure 6 This is a diagram of the HeatVision_YOLO network structure in this application. Figure 7 This is a schematic diagram illustrating the implementation effect of the infrared image enhancement module (IRAE) of this application. Detailed Implementation

[0019] The specification and claims use certain terms to refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The term "comprising" throughout the specification and claims is an open-ended term and should be interpreted as "comprising but not limited to." "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error.

[0020] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0021] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0022] like Figures 1 to 6 As shown, this application provides a method for detecting composite insulators based on multi-scale infrared enhancement, which specifically includes the following steps: Based on the YOLOv11 framework, an infrared composite insulator detection model is constructed to detect composite insulators in infrared images. The construction method of the infrared composite insulator detection model is as follows: The C3k2_CG module is embedded in the shallow and middle layers of the YOLO backbone to enhance the joint modeling of local textures and surrounding context, thereby improving the ability to characterize the slender structure of composite insulators. In the spatial pyramid pooling stage of the YOLO backbone, the SPPF_SE module is used to recalibrate the channels of high-level features and perform multi-scale pooling and splicing to enhance the high-level semantic representation in complex backgrounds. An ASF_YOLO feature fusion unit is embedded between the YOLO backbone and the Neck to extract and fuse multi-scale features such as P3, P4, and P5, thereby improving cross-scale information interaction and positioning accuracy. SimAM nonparametric attention modules are inserted into the feature maps of the YOLO detection head at various scales to significantly enhance the multi-scale features, suppress background interference, and strengthen the target response. In the pre-training stage of the infrared composite insulator detection model, an infrared image enhancement module (IRAE) is introduced to enhance the infrared contrast and structural details of the samples, providing better input for subsequent feature extraction and detection.

[0023] Traditional data augmentation methods are primarily designed for visible light images, and their direct application to infrared images presents numerous limitations. Infrared images reflect the thermal radiation characteristics of objects, rather than the reflected light characteristics of visible light, resulting in fundamental differences in their physical imaging mechanisms. Furthermore, the thermal noise generated by infrared sensors exhibits specific distribution characteristics, significantly different from the noise patterns in visible light images. In addition, infrared imaging is extremely sensitive to changes in ambient temperature; fluctuations in environmental factors significantly affect imaging performance. The contrast distribution of infrared images also differs significantly from that of visible light images, making it difficult to directly transfer and apply traditional contrast augmentation methods designed for visible light. To address the unique imaging mechanism and noise characteristics of infrared images, this application designs a dedicated data augmentation module, IRAE, comprising four core sub-modules: adaptive contrast augmentation, thermal noise simulation, temperature change simulation, and boundary-preserving filtering. This module significantly improves the model's generalization ability and robustness to complex infrared scenes by simulating various interference factors in a real infrared imaging environment.

[0024] (1) Contrast adaptive enhancement module; The contrast of infrared images is affected by many factors such as ambient temperature, target material and atmospheric conditions. This module adopts an adaptive strategy to dynamically adjust the enhancement parameters according to the average brightness value of the image. In order to achieve adaptive contrast enhancement of infrared images, this module uses the mathematical model of formula (1) to process the input image. For dark images (mean < 100), a strong contrast enhancement and brightness enhancement are adopted to highlight the low temperature target features. For bright images (mean ≥ 100), a moderate contrast enhancement and slight brightness adjustment are adopted to avoid overexposure. The contrast is adaptively adjusted according to the local statistical characteristics of the image to improve the distinction between the target and the background. (1) in, For input pixels, and Here, α and β represent the local mean and standard deviation, respectively, and α and β are the tensile parameters.

[0025] (2) Thermal noise simulation module; Infrared sensors generate various types of noise during operation. Gaussian noise simulation of infrared images typically uses an additive white Gaussian noise model, which generates a normally distributed noise matrix with the same size as the original image matrix, a mean of zero, and a variance of one. This matrix is ​​then directly superimposed onto the image. Alternatively, the mean and variance parameters of the Gaussian noise can be flexibly set according to the noise level of different thermal imaging systems to achieve a noise simulation effect that better meets the needs of practical applications. This module uses a composite noise model on the original infrared image... Gaussian noise is superimposed on the data. Gaussian thermal noise simulates random noise generated by thermal motion. The enhanced data can realistically reflect the common noise characteristics of infrared images in actual applications, thereby improving the robustness and generalization ability of the model. The model can better adapt to various noise interferences in actual infrared images. (2) in, With a mean of 0 and a variance of Gaussian noise.

[0026] (3) Temperature change simulation module; Changes in ambient temperature can significantly affect the imaging quality and target visibility of infrared images. This module enhances the environmental adaptability of the model by simulating the imaging effect under different temperature conditions. By linearly adjusting the overall pixel values ​​of the input image, the module simulates the effect of temperature increase or decrease, thereby enhancing the environmental robustness of the model. (3) Where ΔT is the temperature change, and k is the influence coefficient of temperature change on the pixel.

[0027] (4) Boundary-preserving filter module; Infrared images may experience boundary blurring or sharpening during transmission and processing, affecting the clarity of target edge details. This module employs bilateral filtering and unsharpened masking techniques to maintain target boundary clarity while performing blurring or sharpening processing, avoiding damage to the insulator's contour features. and The weighted average preserves the sharpness of target edges during denoising or sharpening, suppresses background noise and blur, and improves the accuracy of target detection and image quality. (4) in, For a spatial Gaussian kernel, The Gaussian kernel for pixel value differences, Ω is the normalization coefficient, and Ω is the neighborhood window.

[0028] The basic YOLOv11n model uses a continuous C3k2 structure in the backbone network to extract multi-scale features. However, in the scenario of slender targets such as infrared composite insulators, traditional convolutions struggle to simultaneously capture local texture details and a broader contextual structure. This application introduces the C3k2_CG module into the shallow and middle layers of the YOLO backbone. While maintaining the lightweight characteristics of the C2f framework, it introduces ContextGuidedBlock with contextual modeling capabilities to enhance the representation of complex backgrounds such as insulator strings and surrounding conductors and fittings, thereby improving the accuracy and robustness of subsequent detection. The C3k2_CG structure is as follows: Figure 1 As shown.

[0029] ContextGuidedBlock (CGB) is a module used to enhance the utilization of contextual information. It mainly addresses the problem of limited receptive field in convolutional neural networks (CNNs) when processing images. By introducing global or local contextual information, it guides the feature extraction process, thereby improving the model's ability to understand the target.

[0030] The structure of the ContextGuidedBlock in the C3k2_CG module of this application is as follows: Figure 2 As shown, after dimensionality reduction of the input features through 1×1 convolution, they are fed into a local feature extraction branch using standard 3×3 channel convolution and a surrounding context extraction branch using dilated channel convolution, respectively. The two outputs are concatenated along the channel dimension and formed into joint features through batch normalization and PReLU nonlinear transformation. Subsequently, a global context extraction unit composed of global average pooling and fully connected layers is used to generate channel attention weights for the joint features, and each channel is adaptively recalibrated to achieve emphasis enhancement of channels related to composite insulators and suppression of background noise channels. Thus, the collaborative fusion of local detail capture, surrounding context modeling and global semantic constraints is completed simultaneously within a single module.

[0031] More specifically, the ContextGuidedBlock module consists of four parts: a local feature extractor, a surrounding context extractor, a joint feature extractor, and a global context extractor. The local feature extractor uses a standard 3×3 convolution to process the feature map after 1×1 dimensionality reduction, focusing on learning local information such as target edges and detailed textures to capture subtle structural changes on the surface of the composite insulator. The surrounding context extractor uses channel convolution with a dilation rate to perform 3×3 dilated convolution on the same number of channels. By expanding the receptive field of the sheet convolution, a wider range of contextual information is introduced, enabling the network to obtain the structural relationship between the conductors, fittings and background areas around the insulator while perceiving local details. The joint feature extractor concatenates the outputs of the local feature extractor and the surrounding context extractor along the channel dimension, and reshapes and smooths the concatenated features through nonlinear transformations such as batch normalization and PReLU to obtain a joint feature representation that simultaneously encodes local details and the surrounding context. The global context extractor performs global average pooling on the joint feature map and concatenates it with a multi-layer fully connected network to generate a set of channel attention weights. It then recalibrates the joint features channel by channel, thereby adaptively enhancing the channels related to the composite insulator target and suppressing background noise channels based on the global distribution of the entire infrared image, thus improving the discriminativeness of the output features of the C3k2_CG module.

[0032] The basic YOLOv11n model uses an SPPF structure to aggregate multi-scale receptive fields in the upper layers of the backbone. However, in low-contrast infrared scenes, the upper-layer features are often mixed with a large number of background responses that are irrelevant to the target, making it difficult to fully highlight the semantic information of the composite insulator itself.

[0033] To address this, this application introduces the SPPF_SE module in the spatial pyramid pooling stage of the YOLO backbone, replacing the original SPPF structure with SPPF_SE. A channel attention modeling step is added before multi-scale pooling and feature concatenation to selectively enhance and suppress high-level features. Its overall structure is as follows: Figure 3 As shown.

[0034] SPPF_SE is a composite structure that combines SPPF (Spatial Pyramid Pooling - Fast) and SE (Squeeze-and-Excitation) modules, aiming to simultaneously enhance the multi-scale expressive power of features and the channel attention mechanism.

[0035] The SELayerV2 module in this application employs a multi-branch compression excitation mechanism. It first performs global average pooling on the input high-level features to obtain channel description vectors. Then, it extracts multi-view channel semantics through four parallel dimensionality-reduced fully connected branches and concatenates them along the channel dimension. Subsequently, it is fed into a fusion fully connected layer and Sigmoid activation to generate normalized channel weights, which are then multiplied element-wise with the original feature map to complete channel recalibration. The high-level features weighted by SELayerV2 then enter the SPPF's 1×1 convolution and multiple 5×5 max pooling and concatenation processes. This achieves multi-scale spatial aggregation while highlighting channels with stronger responses to insulator targets and suppressing background and noise channels, thereby providing more discriminative high-level multi-scale features for subsequent detection heads.

[0036] The calculation process can be expressed by the following formula: (5) Where x represents the input feature map, Sq This represents the description of the multi-branch path obtained from the squeeze operation. ∑Sq This represents the aggregation of descriptions of various branch passages. Ex This represents the channel attention weights obtained from the excitation mapping. F This indicates that the input features are weighted and labeled channel by channel.

[0037] The traditional Feature Pyramid Network (FPN) of the basic YOLOv11n model suffers from detail loss in PCB detection. This application embeds an ASF_YOLO feature fusion unit between the YOLO backbone and the Neck, specifically referencing the classic weighted bidirectional feature pyramid network ASF_YOLO (Attentional Scale Sequence Fusion based YOLO). Building upon the traditional YOLO detection framework, it integrates an attention-driven multi-scale feature fusion mechanism to effectively integrate spatial and semantic features of targets at different scales, thereby significantly improving detection accuracy while maintaining real-time performance. The ASF_YOLO network structure is as follows: Figure 4 As shown.

[0038] The specific method for extracting multi-scale fusion features using the ASF_YOLO feature fusion unit is as follows: The SSFF submodule uses Zoom_cat and Concat operations to perform scale alignment and multi-scale fusion of features in layers P3, P4, and P5; The TFE submodule uses the ScalSeq operation to decompose and enhance the fused features at three scales: large, medium, and small. The CPAM submodule introduces cross-scale attention at the Add position to integrate information from SSFF and TFE.

[0039] This application's ASF_YOLO framework combines global semantic information from multi-scale images. The SSFF (Scale-Sequential Feature Fusion) module effectively fuses feature maps from layers P3, P4, and P5 to capture spatial information from different locations, covering target features of various sizes and shapes. It also normalizes the feature maps from different layers to ensure a uniform spatial resolution. Finally, the processed multi-scale feature maps are stacked as input for subsequent convolution operations. The feature maps from different layers are convolved with a series of Gaussian kernels with increasing standard deviations, as expressed by the formula: (6) (7) Where f(w,h) represents a two-dimensional input image with width w and height h, respectively. It consists of a series of convolutions using a two-bit Gaussian filter. The smoothing process generates σ, which is the standard deviation scaling parameter of the two-dimensional Gaussian filter used for convolution.

[0040] The fusion mechanism of the traditional Feature Pyramid Network (FPN) ignores the rich details of larger feature layers, only upsampling small feature maps and splitting them to add to the previous layer. It ignores the scale normalization problem. The TFE module can obtain complete and detailed feature information by segmenting large, medium and small features, adding larger feature maps and amplifying the features.

[0041] The TFE module processes feature maps of different sizes by adjusting the number of feature channels, using a hybrid pooling structure for downsampling, using transposed convolution for upsampling, and employing an attention weight allocation mechanism. It then fuses feature maps of different sizes and concatenates them along the channel dimension, as shown in the following formula: (8) in, The feature map output by the TFE module. , , These represent feature maps of large, medium, and small sizes, respectively. resolution and The same, and the number of channels is 3 times.

[0042] ASF-YOLO utilizes the CPAM module to integrate detailed information and multi-scale feature information from SSFF and TFE. The CPAM module consists of a channel attention network (Input1) that receives TFE input, an input position attention network (Input2) that receives channel attention network and SSFF output superimposed. To effectively capture cross-channel interactions and avoid the problems of fully connected layers, an adaptive attention weight allocation mechanism based on 1D convolution is introduced. Channel attention is achieved by considering the local cross-channel interactions of each channel and its K nearest neighbors, and an adaptive relationship between the channel dimension C and the convolution kernel K is established, as follows: (9) (10) With γ set to 2 and b set to 1, the mapping relationship above shows that high-value channels have longer exchange times, while low-value channels have shorter exchange times. Therefore, the channel attention weight allocation mechanism can more deeply mine the features of multiple channels. Here, γ and b are scaling parameters used to control the ratio of the convolution kernel size k to the channel dimension C. It is an odd number of nearest neighbors.

[0043] The SSFF module introduces a synergistic effect of positional attention weight allocation and channel attention weight allocation mechanisms. Based on average pooling (pw) and max pooling (ph), it performs feature distribution analysis and weight allocation to preserve the spatial structure information of the feature map, achieving effective aggregation of spatial encoding information. Its mathematical expression is as follows: (11) (12) Here, E(w,j) and E(i,h) are the input feature maps at position (i,j).

[0044] When generating position attention coordinates, the CPAM module applies a concatenation operation to the horizontal and vertical axes to obtain the position attention coordinates, and then uses a split operation to pair the position-dependent feature maps, as shown in the following formula: (13) (14) (15) Wherein, P( , ) is the output of the position attention coordinates. and These are the width and height of the segmented output, respectively.

[0045] Finally, the output of the CPAM module is defined by the following formula: (16) Here, E represents the weights of channel attention and position attention.

[0046] In the task of detecting defects in infrared composite insulators of transmission lines, due to the small difference in thermal radiation between the target and the background at low temperatures at night, and the presence of complex power equipment interference, traditional neural networks struggle to effectively focus on key defect areas. To address this issue, a parameter-free three-dimensional attention weight allocation mechanism, SimAM (Simple, Parameter-Free Attention Module), is introduced, with the structure as follows: Figure 4 As shown, by learning the similarity between pixels in the feature map, attention is focused on more important and target-related locations, while irrelevant background information is suppressed. Unlike traditional channel or spatial attention weight allocation mechanisms such as SE, ECA, and CBAM, SimAM does not require the introduction of additional parameters and generates attention weights by calculating the energy function of each neuron and its neighboring neurons.

[0047] In infrared composite insulator detection, SimAM can adaptively enhance the feature response of temperature anomaly areas on the insulator surface while suppressing background noise and interference from irrelevant areas. This mechanism effectively improves the network's ability to perceive subtle temperature changes in infrared images through parameterless energy function calculation, and assigns more accurate three-dimensional attention weights to the features in the feature map, thereby significantly improving the accuracy and robustness of composite insulator defect detection.

[0048] Inspired by the phenomenon of spatial inhibition in visual neuroscience, SimAM assesses the importance of neurons by measuring their linear separability. To this end, the following energy function is defined for each neuron: (17) (18) (19) In the formula: Let λ be the minimum energy function, t be the target neuron, and λ be the regularization coefficient. and , , are the mean values ​​of all neurons in that channel, and M is the number of neurons in each channel.

[0049] When the energy of a neuron is lower, it indicates that the neuron is significantly different from its neighboring neurons, and its importance increases accordingly. The result after feature enhancement is shown in formula (20): (20) In the formula: X is the input feature, It is the output characteristic, and E is the sum of all energies.

[0050] After the detection head outputs features at each scale of the original YOLOV11n model, a lightweight, parameter-free attention weight allocation mechanism, SimAM, is introduced. This does not increase the number of model parameters or computational complexity. SimAM is applied in parallel at the P3, P4, and P5 scales to achieve comprehensive attention enhancement for small targets. It is combined with the SSFF (Zoom_cat) module of ASF_YOLO to build an efficient multi-module fusion architecture, further improving the accuracy of infrared composite insulator target detection.

[0051] Based on the above analysis, this application proposes an improved model, HeatVision_YOLO, such as... Figure 6 As shown, the IRAE is placed outside the network as a data-side preprocessing module, and the overall architecture adopts a classic three-layer design of "backbone network - neck network - detection head": The backbone network consists of Conv convolutional layers, C3k2_CG modules embedded with ContextGuidedBlock, deep C3k2 modules, SPPF_SE spatial pyramid pooling modules with SELayerV2 channel attention, and C2PSA adaptive receptive field modules, which are used to enhance local texture and context modeling in shallow and middle layers, and enhance channel selection and multi-scale semantic representation in high layers. The neck network integrates ASF_YOLO features, utilizes SSFF to achieve efficient alignment and fusion of multi-scale features, leverages TFE to complete serialized feature rearrangement and enhancement, and performs weighted feature fusion through CPAM; The detection head embeds the SimAM parameterless attention weight allocation mechanism for feature enhancement at three different scales: P3, P4, and P5. Finally, it outputs multi-scale detection results through three Detect modules.

[0052] This architecture achieves organic synergy between multimodal feature integration, spatial information enhancement, and attention weight allocation, enabling comprehensive coverage and accurate detection of small, medium, and large targets. It is particularly suitable for target detection tasks in complex scenarios such as infrared images of power equipment.

[0053] Example 1 This study used images captured by a municipal power supply bureau during inspections using a FLIR infrared thermal imager, collecting a total of 514 images as a pre-training dataset. To enhance the richness of the dataset, data augmentation methods were employed to expand its size, reduce the dependence of the composite insulator identification model on specific image attributes, mitigate overfitting in the training model, and improve its robustness. A dedicated data augmentation module was designed, comprising four core sub-modules: contrast enhancement, thermal noise simulation, temperature change simulation, and boundary-preserving filtering. This module significantly improved the model's adaptability to complex infrared scenes by simulating various changes in the real infrared imaging environment. After data augmentation, the pre-training dataset expanded six times to a total of 3084 images. Figure 7 The comparison example shows the results after data augmentation.

[0054] This embodiment uses PyTorch 2.1.8 to build a deep learning model, and the parameters are shown in Table 1.

[0055] Table 1. Experimental Platform Configuration Parameters Precision (P), recall (R), and mean average precision (mAP) with IoU values ​​of 0.5 and 0.5-0.95 were selected as evaluation indicators for the model's accuracy in detecting composite insulators. Among them, mAP is the key evaluation indicator for the detection accuracy of this application. The closer P, R, and mAP are to 1, the higher the model's detection accuracy. The calculation formulas for the three are shown in (20)-(22). (twenty one) (twenty two) (twenty three) In the formula, TP (True Positive) represents the number of correctly identified targets, FP (False Positive) represents the number of targets misidentified, FN (False Negative) represents the actual number of undetected targets, and AP (Average Precision) serves as the core evaluation metric, comprehensively measuring detection accuracy by calculating the area under the PR curve. To comprehensively evaluate the model's detection accuracy, computational efficiency, and deployment feasibility, a multi-dimensional evaluation system is also constructed using model parameter count (Params) and computational cost (GFLOPs) as evaluation metrics.

[0056] First, a comparative data augmentation experiment was conducted. Four augmentation sub-modules were integrated under a probabilistic scheduling strategy to form a complete infrared image augmentation pipeline. Each sub-module employed independent probabilistic control to ensure the diversity and randomness of the augmentation, effectively mitigating overfitting. The IRAE module consists of four parts: contrast enhancement, thermal noise simulation, temperature change simulation, and boundary-preserving filtering. Taking an infrared composite insulator image with a complex background as an example (e.g....) Figure 7 ), Figure 7 (a) is the original image. Figure 7 (f) shows the overall effect after the four modules work together. The contrast enhancement significantly increases the brightness difference between the insulator and the background, alleviating the boundary blurring caused by low contrast. Figure 7 (b) Thermal noise simulation replicates sensor and environmental disturbances in real-world scenarios, reducing false detections in complex backgrounds. Figure 7 (c) Temperature changes characterize the differences in heat distribution across environments, improving cross-domain generalization. Figure 7 (d) Boundary-preserving filtering retains the slender outline of the composite insulator while maintaining the background texture, improving positioning accuracy. Figure 7 As shown in (e), the four modules work together to improve accuracy and reduce false detections by combining the multi-scale fusion of ASF_YOLO and the attention mechanism of SimAM. While maintaining a relatively stable recall rate, they meet the engineering application requirements of high precision and strong robustness in infrared composite insulator scenarios.

[0057] To verify the effectiveness of each improved module, ablation experiments were first conducted. The IRAE module, C3k2_CG module, SPPF_SE module, ASF_YOLO module, and the parameterless attention weight allocation mechanism SimAM module were introduced into YOLOv11n for comparative experiments. Each group of experiments used the same training method and hyperparameter settings.

[0058] Table 2 shows that, based on the YOLOv11n baseline model, introducing only IRAE improved the precision from 85.8% to 89.9%, while mAP remained basically the same. This indicates that infrared enhancement mainly improves feature discriminability and precision, but sacrifices some recall. Further adding ASF_YOLO significantly improved the recall from 79.2% to 86.9%, with mAP@0.5 and mAP@0.5:0.95 increasing by 2.4% and 3.9% respectively. This demonstrates that multi-scale feature fusion effectively recovers and improves the recall and overall detection quality of targets at various scales. Continuing to overlay SimAM further improved both precision and mAP. The indicators showed a slight improvement, indicating that the parameterless attention of the detection head can optimize feature weight allocation without increasing the computational load. Based on this, the introduction of backbone enhancement modules such as C3k2_CG to form the final HeatVision_YOLO improved precision, recall, mAP@0.5 and mAP@0.5:0.95 to 87.9%, 86.6%, 94.1% and 78.1% respectively, with the four indicators being the best overall.

[0059] The overall ablation results show that IRAE focuses on enhancing the quality of infrared features, and ASF_YOLO and SimAM together improve multi-scale fusion and attention allocation. Combined with the improved backbone structure, the detection accuracy and reliability are significantly improved, reducing the pressure of subsequent manual re-inspection and verifying the effectiveness of the proposed method.

[0060] Table 2 Ablation Experiment Results Table 2 Ablation Study Results Under the same data partitioning, training strategy, and input scale settings, mainstream lightweight detection models such as YOLOv5n, YOLOv8n, YOLOv10n, and YOLOv11n were selected as comparison baselines. The detection results are shown in Table 3. HeatVision_YOLO achieved the best performance in three out of the four metrics: compared to the baseline model YOLOv8n with the best overall performance, its mAP@0.5 improved from 92.6% to 94.1%, an improvement of 1.5%. mAP@0.5–0.95 improved from 74.9% to 78.1%, an improvement of 3.2%. Recall improved from 82.8% to 86.6. The advantages in cross-threshold mAP and recall indicate fewer missed detections and more stable localization. Compared to the main baseline YOLOv11n at the same scale, HeatVision_YOLO improved by 2.1, 2.9, 3.8, and 5.4 percentage points in Precision, Recall, mAP@0.5, and mAP@0.5:0.95, respectively. In summary, the synergistic effect of IRAE infrared enhancement, ASF_YOLO multi-scale feature fusion, C3k2_CG and SPPF_SE backbone enhancement, and SimAM detector head attention effectively suppresses low contrast, thermal noise, and complex background interference, significantly improving the robustness and generalization ability of the model.

[0061] Experiments show that, compared with the traditional YOLOv11n, HeatVision_YOLO achieves higher accuracy in infrared small target detection by sacrificing a small amount of speed, significantly reducing the false detection rate, improving positioning quality, and performing more stably under cross-scene conditions. Overall, it improves the accuracy and robustness of infrared image detection of composite insulators, providing a high-precision and robust solution for engineering applications.

[0062] Table 3 Detection results under different methods The embodiments of this application have been described in detail above with reference to the accompanying drawings. However, this application is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of this application.

Claims

1. A method for detecting composite insulators based on multi-scale infrared enhancement, characterized in that: Based on the YOLOv11 framework, an infrared composite insulator detection model was constructed to detect composite insulators in infrared images. The specific construction of the infrared composite insulator detection model is as follows: The C3k2_CG module is embedded in the shallow and middle layers of the YOLO backbone to enhance the joint modeling of local textures and surrounding context; In the spatial pyramid pooling stage of the YOLO backbone, the SPPF_SE module is used to recalibrate the channels of the high-level features and perform multi-scale pooling and stitching. An ASF_YOLO feature fusion unit is embedded between the YOLO backbone and the Neck to extract and fuse multi-scale features. SimAM nonparametric attention modules are inserted into the feature maps at each scale of the YOLO detector head to enhance the multi-scale features; In the pre-training stage of the infrared composite insulator detection model, an infrared image enhancement module (IRAE) is introduced to enhance the infrared contrast and structural details of the samples.

2. The composite insulator detection method based on multi-scale infrared enhancement according to claim 1, characterized in that: The infrared image enhancement module (IRAE) includes a contrast adaptive enhancement module, a thermal noise simulation module, a temperature change simulation module, and a boundary preservation filtering module. The functions of each module are as follows: The contrast adaptive enhancement module employs an adaptive strategy to dynamically adjust enhancement parameters based on the average brightness value of the image in order to achieve adaptive contrast enhancement of infrared images. The thermal noise simulation module uses a composite noise model, which superimposes Gaussian noise onto the original infrared image. Gaussian thermal noise simulates random noise generated by thermal motion. The temperature change simulation module simulates the imaging effect under different temperature conditions; The boundary preservation filtering module employs bilateral filtering and unsharpening masking techniques to maintain the clarity of the target boundary while performing blurring or sharpening processing.

3. The composite insulator detection method based on multi-scale infrared enhancement according to claim 2, characterized in that: The contrast adaptive enhancement module processes the input image using a mathematical model. If the image is a dark image with a mean <100, the stretching coefficient α associated with contrast enhancement is [1.3, 1.8], and the stretching coefficient β associated with brightness enhancement is [10, 30]. If the image is a bright image with a mean ≥100, the stretching coefficient α associated with contrast enhancement is [1.1, 1.5], and the stretching coefficient β associated with brightness enhancement is [0, 15]. The mathematical model is as follows: in, For input pixels, and Here, α and β represent the local mean and standard deviation, respectively, and the tensile parameters are α and β. The thermal noise simulation module employs a composite noise model in the original infrared image. Gaussian noise is added on top of the existing structure, as shown in the following formula: in, With a mean of 0 and a variance of Gaussian noise; The temperature change simulation module simulates the effect of temperature increase or decrease by linearly adjusting the overall pixel values ​​of the input image, as shown in the following formula: Where ΔT is the temperature change, and k is the influence coefficient of temperature change on the pixel; The boundary preservation filter module, through and The weighted average preserves the sharpness of the target edges during denoising or sharpening, as shown in the following formula: in, For input pixels, For a spatial Gaussian kernel, The Gaussian kernel for pixel value differences, Ω is the normalization coefficient, and Ω is the neighborhood window.

4. The composite insulator detection method based on multi-scale infrared enhancement according to claim 2, characterized in that: The C3k2_CG module replaces the traditional convolutional bottleneck in the bottleneck branch of the original C2f structure with ContextGuidedBlock. ContextGuidedBlock concatenates local features with contextual features, and then introduces global channel attention for joint modeling. The specific method is as follows: After dimensionality reduction of the input features using 1×1 convolution, they are fed into the local feature extraction branch using standard 3×3 channel convolution and the surrounding context extraction branch using dilated channel convolution, respectively. The two outputs are spliced ​​together along the channel dimension and then combined using batch normalization and PReLU nonlinear transformation to form a joint feature. A global context extraction unit consisting of global average pooling and fully connected layers is used to generate channel attention weights for joint features, and adaptive recalibration is performed on each channel.

5. The composite insulator detection method based on multi-scale infrared enhancement according to claim 2, characterized in that: The SPPF_SE module utilizes its SELayerV2 submodule to perform channel recalibration on high-level features and execute multi-scale pooling and concatenation, as follows: Global adaptive pooling and channel weighting are applied to the input high-level features; The input feature map is first subjected to global average pooling to compress the spatial dimension into channel description vectors. Then, the channel information is nonlinearly transformed through a multi-branch fully connected layer and spliced ​​and aggregated in the channel dimension. An excitation operation is performed on the aggregated channel description vectors. Normalized channel attention weights are generated through a fully connected layer and a sigmoid activation, and then multiplied element-wise with the input feature map to complete channel recalibration. The recalibrated features are then fed into the residual convolution branch F(·), and finally added to the input feature x using residuals. The calculation process is expressed by the following formula: Where x represents the input feature map, Sq This represents the description of the multi-branch path obtained from the squeeze operation. ∑Sq This represents the aggregation of descriptions of various branch passages. Ex This represents the channel attention weights obtained from the excitation mapping. F This indicates that the input features are weighted and labeled channel by channel.

6. The composite insulator detection method based on multi-scale infrared enhancement according to claim 1, characterized in that: The specific method for extracting multi-scale fusion features using the ASF_YOLO feature fusion unit is as follows: The SSFF submodule uses Zoom_cat and Concat operations to perform scale alignment and multi-scale fusion of features in layers P3, P4 and P5. The TFE submodule utilizes the ScalSeq operation to decompose and enhance the fused features at three scales: large, medium, and small. The CPAM submodule introduces cross-scale attention at the Add position to integrate information from SSFF and TFE. The SimAM parameterless attention module embeds the SimAM parameterless attention weight allocation mechanism for feature enhancement at three different scales: P3, P4, and P5. Finally, it outputs multi-scale detection results through three Detect modules.

7. The composite insulator detection method based on multi-scale infrared enhancement according to claim 6, characterized in that: The SSFF module effectively fuses feature maps from layers P3, P4, and P5, capturing different spatial information and covering target features of various sizes and shapes. It also normalizes the feature maps from different layers to ensure a uniform spatial resolution. Finally, the processed multi-scale feature maps are stacked as input for subsequent convolution operations. The feature maps from different layers are convolved with a series of Gaussian kernels with increasing standard deviations, as expressed by the formula: Where f(w,h) represents a two-dimensional input image with width w and height h, respectively. It consists of a series of convolutions using a two-bit Gaussian filter. The smoothing process generates σ, which is the standard deviation scaling parameter of the two-dimensional Gaussian filter used for convolution.

8. The composite insulator detection method based on multi-scale infrared enhancement according to claim 6, characterized in that: The TFE module processes feature maps of different sizes by adjusting the number of feature channels, using a hybrid pooling structure for downsampling, using transposed convolution for upsampling, and employing an attention weight allocation mechanism. It then fuses feature maps of different sizes and concatenates them along the channel dimension, as shown in the following formula: in, The feature map output by the TFE module. , , These represent feature maps of large, medium, and small sizes, respectively. resolution and The same, and the number of channels is Three times that.

9. The composite insulator detection method based on multi-scale infrared enhancement according to claim 6, characterized in that: The CPAM module consists of a channel attention network Input1 that receives TFE input, an input position attention network Input2 that receives channel attention networks and SSFF outputs, and introduces an adaptive attention weight allocation mechanism based on 1D convolution. Channel attention is achieved by considering the local cross-channel interactions of each channel and its K nearest neighbors, and an adaptive relationship between the channel dimension C and the convolution kernel K is established, as follows: γ is set to 2, and b is set to 1. γ and b are scaling parameters used to control the ratio of the convolution kernel size k to the channel dimension C. It is an odd number of nearest neighbors.

10. The composite insulator detection method based on multi-scale infrared enhancement according to claim 9, characterized in that: The SSFF module introduces a synergistic effect of positional attention weight allocation and channel attention weight allocation mechanisms. Based on average pooling (pw) and max pooling (ph), it performs feature distribution analysis and weight allocation to preserve the spatial structure information of the feature map, achieving effective aggregation of spatial encoding information. Its mathematical expression is as follows: Where E(w,j) and E(i,h) are the input feature maps at position (i,j); When generating position attention coordinates, the CPAM module applies a concatenation operation to the horizontal and vertical axes to obtain the position attention coordinates, and then uses a split operation to pair the position-dependent feature maps, as shown in the following formula: Wherein, P( , ) is the output of the position attention coordinates. and These are the width and height of the segmented output, respectively; Finally, the output of the CPAM module is defined by the following formula: Here, E represents the weights of channel attention and position attention.

Images