Power equipment infrared image detection method based on improved YOLO11 network

Abstract

The application discloses an improved YOLO11 network-based power equipment infrared image detection method, which comprises the steps of infrared image acquisition and preprocessing, improved YOLO11 network construction, network training and the like. By integrating an improved feature calibration (AFCM) module in the backbone network of the YOLO11 network, semantic and spatial features are decoupled; by deploying a dynamic content guidance fusion (DCG-Fusion) module in the neck network, adaptive alignment of different scale features is realized; and by introducing a kernel point distance loss (IPD-Loss) in the training stage, the focusing on the target core area is strengthened by constructing an internal auxiliary frame. The application can not only cope with extreme environmental interference, but also has the ability to process multi-scale and multi-target detection, thereby providing reliable technical support for intelligent inspection and equipment state evaluation of a power system.

Description

Technical Field

[0001] This invention relates to the technical field of computer vision and power system automation, and in particular to a method for infrared image detection of power equipment based on an improved YOLO11 network. Background Technology

[0002] In the field of power system automation monitoring, infrared imaging technology, due to its unique imaging principle, can provide equipment operation information beyond the visible light environment, making it an important data source for all-weather status monitoring of power facilities. However, in complex power operating environments, how to achieve efficient, accurate, and automated identification of power equipment in infrared images remains a core challenge in current intelligent operation and maintenance.

[0003] Currently, although computer vision and deep learning technologies have made significant progress in the field of general object detection, the results of directly applying these algorithms to power infrared image detection are often unsatisfactory. First, infrared images generally suffer from low contrast and blurred edge textures. Complex and variable heat radiation sources in the background often overlap or interfere with the features of the equipment being inspected, making it difficult for models to stably extract semantic information from the target. Second, power system facilities have a wide range of physical dimensions. Under the same imaging viewpoint, the pixel proportions of different targets in the image vary significantly. Existing multi-scale feature fusion mechanisms are prone to problems such as missing small targets or incomplete perception of large targets when processing such targets. Furthermore, power components have diverse shapes and numerous non-standard geometric configurations. Traditional localization regression algorithms often suffer from low convergence accuracy when dealing with targets with such specific geometric distributions.

[0004] In summary, existing technologies for processing power equipment infrared images lack a comprehensive solution that can simultaneously address background interference resistance, multi-scale feature adaptation, and high-precision geometric positioning, making it difficult to meet the demands of high-reliability inspections. This invention aims to provide a more intelligent and robust infrared image detection solution for power equipment to overcome the shortcomings of existing technologies in complex environments. Summary of the Invention

[0005] The purpose of this invention is to provide an infrared image detection method for power equipment based on an improved YOLO11 network. This method can achieve efficient and accurate positioning and identification of multi-scale and irregularly shaped power equipment under complex power background conditions, improve the intelligence level of power automated inspection, and provide reliable data support for power equipment condition assessment and operation and maintenance decisions.

[0006] To achieve the above objectives, the technical solution provided by this invention is as follows: a method for detecting infrared images of power equipment based on an improved YOLO11 network. The improved YOLO11 network improves the backbone network, neck network, and loss function of the original network. Specifically, the improvement to the backbone network involves constructing an improved feature calibration module, called the AFCM (Advanced Feature Complementary Mapping) module. The AFCM module extracts heterogeneous features through decoupled semantic and spatial branches, and utilizes a dual-criteria weight extraction mechanism based on global average pooling and global max pooling to suppress infrared noise and enhance key equipment features. The improvement to the neck network involves constructing a dynamic content guided fusion module, called the DCG-Fusion (Dynamic Content Guided Fusion) module, replacing the original feature stitching structure. The DCG-Fusion module extracts mutation information by calculating the difference features between shallow and deep layers, and uses parallel dilated convolution sequences to generate multi-scale gating weights, achieving efficient perception of power targets of different sizes. The improvement to the loss function involves constructing a kernel point distance loss, called IPD-Loss (Inner... Point Distance Loss (IPD-Loss) is a method that uses a scale factor to construct an internal auxiliary box to focus on the core area of ​​the target and uses the corner Euclidean distance mapping between the predicted auxiliary box and the real auxiliary box to improve the shape fitting ability of irregularly shaped power equipment.

[0007] The specific implementation of the infrared image detection method for power equipment includes the following steps:

[0008] Step 1: Collect raw infrared images of power equipment, and after image enhancement and annotation, construct a power infrared detection dataset and perform image preprocessing to meet the requirements of the improved YOLO11 network input.

[0009] Step 2: Input the power infrared detection dataset into the improved YOLO11 network for training. During training, the images in the training set are decoupled and extracted by the AFCM module, and multi-scale dynamic fusion is performed in the neck network by the DCG-Fusion module. During backpropagation, the regression loss is calculated using the IPD-Loss and the network parameters are updated to train the improved YOLO11 network with the best performance.

[0010] Step 3: Input the infrared image of the power equipment to be detected into the improved YOLO11 network with the best performance obtained from training, and output the power equipment detection results.

[0011] Furthermore, step 1 includes the following steps:

[0012] Step 1.1, Power Infrared Data Acquisition: Use an infrared thermal imager to capture images of the power operation scene and obtain the original infrared image sequence containing the power equipment;

[0013] Step 1.2, Infrared image quality enhancement: Adaptive histogram equalization or bilateral filtering algorithm is used to enhance the contrast and edge details of all infrared images in the original infrared image sequence;

[0014] Step 1.3, Labeling and Sample Construction: Label the power equipment in the enhanced infrared image, generate a label file containing the target category and location coordinates, and construct a power infrared detection dataset;

[0015] Step 1.4, Multi-scale mapping and normalization: The images and corresponding labels in the power infrared detection dataset are uniformly adjusted to a preset pixel size, and pixel value normalization is performed to make them meet the requirements of the improved YOLO11 network input.

[0016] Furthermore, the specific implementation logic of the AFCM module is as follows:

[0017] Heterogeneous feature channel segmentation: The input feature map X is segmented into a first feature branch and a second feature branch in the channel dimension by a ratio α. By decoupling the category semantics and spatial location information in the initial stage, the decoupled first feature branch is used as the semantic branch and the decoupled second feature branch is used as the spatial branch, so as to improve the network's ability to retain low-level spatial information.

[0018] Heterogeneous branch feature transformation: Standard convolution is performed on the semantic branches to obtain semantic features X containing rich category attributes. C The spatial branches are subjected to point-by-point convolution to obtain spatial features X that preserve the geometric topology. S ;

[0019] A dual-criteria weight extraction mechanism based on global average pooling and global max pooling: based on the semantic feature X C With spatial features X S The channel weights w1 and spatial weights w2 used for complementary calibration are generated, and their calculation formulas are as follows:

[0020] ;

[0021] ;

[0022] In the formula, σ() represents the Sigmoid activation function, MLP() represents the weight mapping operation of the multilayer perceptron, GAP() and GMP() represent global average pooling and global max pooling operations, respectively, and Conv() represents the convolution operation; by introducing global max pooling and global average pooling in parallel, the ability to perceive significant hotspot contours of power equipment in infrared images is enhanced.

[0023] Complementary mapping and feature aggregation: using the w1 to pair spatial features X S Calibration is performed using w2 on semantic feature X. C The calibration is performed, and the calibrated two feature streams are concatenated and aggregated to obtain the output feature X. FCM The calculation formula is as follows:

[0024] ;

[0025] In the formula, Concat() represents the channel concatenation operation.

[0026] Furthermore, all bottleneck units within the C3k2 module in the four stages from Stage 2 to Stage 5 of the backbone network are replaced with the AFCM module to achieve deep complementarity and alignment of spatial geometric features and semantic attribute features throughout the feature extraction process.

[0027] Furthermore, the specific implementation logic of the DCG-Fusion module is as follows:

[0028] Deep features x and shallow features y are mapped to the same dimension, and their difference feature term (xy) is calculated to explicitly extract abrupt changes in target edges in infrared images. The difference feature term is then modeled in a multi-scale space using a parallel dilated convolution sequence with multiple dilation rate branches. Pixel-level dynamic gating weights G are generated via an activation function, and their calculation formula is as follows:

[0029] ;

[0030] In the formula, DConv seq () represents a parallel dilated convolution sequence operation, used to adaptively adjust the feature receptive field to match power equipment of different scales; finally, the pixel-level dynamic gating weights G are used to perform complementary weighted fusion of deep and shallow features to obtain the output feature map Y, the calculation formula of which is:

[0031] .

[0032] Furthermore, the DCG-Fusion module is deployed at the intersection of deep and shallow features in the neck network, and dynamic content-guided fusion is used to solve the problems of large target scale span and noise interference in infrared images of power equipment.

[0033] Furthermore, the specific implementation logic of IPD-Loss is as follows:

[0034] Auxiliary box scaling: A scale factor r is introduced to synchronously scale the width and height of the predicted box and the ground truth box to construct an internal auxiliary box that focuses on the core region of the target.

[0035] Auxiliary point distance calculation: Calculate the squared Euclidean distance between corresponding corner points of the inner auxiliary boxes. The calculation formula is as follows:

[0036] ;

[0037] In the formula, and These represent the coordinates of the top-left corner (n=1) and bottom-right corner (n=2) of the predicted bounding box and the ground truth bounding box, respectively.

[0038] Loss calculation: combining the intersection-union ratio (IoU) of the auxiliary bounding box. inner By mapping the point distance term, we obtain the regression loss function L. IPD :

[0039] ;

[0040] In the formula, IoU inner This represents the intersection-over-union ratio (IoU) between the scaled predicted bounding box and the ground truth bounding box. and , respectively represent the squared distance between corresponding corner points, w and h represent the original image size; the IPD-Loss enhancement network is used to improve the shape fitting ability of power equipment with different aspect ratios and multi-scale, and to suppress the influence of infrared image noise on positioning.

[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0042] 1. This invention significantly improves anti-interference capabilities against complex backgrounds. By introducing feature decoupling extraction, the network can automatically suppress noise in infrared images. Compared to conventional models, this solution can more accurately pinpoint the outline of power equipment, solving the problem of target features being easily interfered with by complex backgrounds.

[0043] 2. This invention achieves efficient sensing of power equipment of different sizes. This scheme generates pixel-level dynamic gating weights by calculating differential feature terms and adaptively adjusts the sensing range. Conventional methods are prone to missed detections when dealing with targets with huge scale spans. This invention, however, can automatically align deep and shallow features, ensuring extremely high detection accuracy for power equipment of various sizes at different shooting distances.

[0044] 3. This invention improves the positioning accuracy of irregularly shaped power equipment. By constructing a kernel point distance loss, this scheme enhances the focus on the core area of ​​the target and utilizes corner distance constraints in the regression process. Conventional positioning algorithms are prone to positioning offsets or convergence difficulties when dealing with targets with extreme aspect ratios. This invention, by introducing a scale factor, significantly improves the network's ability to fit targets of specific shapes, making the positioning results more stable. Attached Figure Description

[0045] Figure 1 This diagram illustrates the structure of an improved YOLO11 network. The network comprises a backbone, a neck, and a head. The backbone integrates an improved feature calibration (AFCM) module; the neck deploys a dynamic content-guided fusion (DCG-Fusion) module; and the head, connected to the end of the network, outputs the final detection results. The remaining modules in the diagram represent their corresponding modules in the original YOLO11 network.

[0046] Figure 2 This is a flowchart illustrating the method of the present invention. The process starts from the initial setup and proceeds sequentially through steps such as data preparation, network training, and target detection, ultimately achieving the training and application of the optimal network. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0048] This embodiment discloses a method for infrared image detection of power equipment based on an improved YOLO11 network, such as... Figure 1As shown, the improved YOLO11 network improves the backbone network, neck network, and loss function of the original network. The improvement to the backbone network involves constructing an improved feature calibration module, called the AFCM (Advanced Feature Complementary Mapping) module. The AFCM module extracts heterogeneous features through decoupled semantic and spatial branches, and utilizes a dual-criteria weight extraction mechanism based on global average pooling and global max pooling to suppress infrared noise and enhance key equipment features. The improvement to the neck network involves constructing a dynamic content guided fusion module, called the DCG-Fusion (Dynamic Content Guided Fusion) module, replacing the original feature splicing structure. The DCG-Fusion module extracts mutation information by calculating the difference features between deep and shallow layers, and uses parallel dilated convolutional sequences to generate multi-scale gating weights, achieving efficient perception of power targets of different sizes. The improvement to the loss function involves constructing a kernel point distance loss, called IPD-Loss (Inner Point Distance). The IPD-Loss constructs an internal auxiliary box through a scale factor to focus on the core area of ​​the target, and uses the corner Euclidean distance mapping between the predicted auxiliary box and the real auxiliary box to improve the shape fitting ability of irregular power equipment.

[0049] like Figure 2 As shown, the specific implementation of the infrared image detection method for power equipment includes the following steps:

[0050] Step 1: Acquire raw infrared images of power equipment, construct a power infrared detection dataset after image enhancement and annotation, and perform image preprocessing to meet the requirements of the improved YOLO11 network input. This includes the following steps:

[0051] Step 1.1, Power Infrared Data Acquisition: Use an infrared thermal imager to capture images of the power operation scene and obtain the original infrared image sequence containing the power equipment;

[0052] Step 1.2, Infrared image quality enhancement: Adaptive histogram equalization or bilateral filtering algorithm is used to enhance the contrast and edge details of all infrared images in the original infrared image sequence;

[0053] Step 1.3, Labeling and Sample Construction: Label the power equipment in the enhanced infrared image, generate a label file containing the target category and location coordinates, and construct a power infrared detection dataset;

[0054] Step 1.4, Multi-scale mapping and normalization: The images and corresponding labels in the power infrared detection dataset are uniformly adjusted to a preset pixel size, and pixel value normalization is performed to make them meet the requirements of the improved YOLO11 network input.

[0055] Step 2: Input the power infrared detection dataset into the improved YOLO11 network for training. During training, the images in the training set are decoupled and extracted by the AFCM module, and multi-scale dynamic fusion is performed in the neck network by the DCG-Fusion module. During backpropagation, the regression loss is calculated using the IPD-Loss and the network parameters are updated to obtain the improved YOLO11 network with the best performance.

[0056] In a specific example, the implementation logic of the AFCM module is as follows:

[0057] Heterogeneous feature channel segmentation: The input feature map X is segmented into a first feature branch and a second feature branch in the channel dimension by a ratio α. By decoupling the category semantics and spatial location information in the initial stage, the decoupled first feature branch is used as the semantic branch and the decoupled second feature branch is used as the spatial branch, so as to improve the network's ability to retain low-level spatial information.

[0058] Heterogeneous branch feature transformation: Standard convolution is performed on the semantic branches to obtain semantic features X containing rich category attributes. C The spatial branches are subjected to point-by-point convolution to obtain spatial features X that preserve the geometric topology. S ;

[0059] A dual-criteria weight extraction mechanism based on global average pooling and global max pooling: based on the semantic feature X C With spatial features X S The channel weights w1 and spatial weights w2 used for complementary calibration are generated, and their calculation formulas are as follows:

[0060] ;

[0061] ;

[0062] In the formula, σ() represents the Sigmoid activation function, MLP() represents the weight mapping operation of the multilayer perceptron, GAP() and GMP() represent global average pooling and global max pooling operations, respectively, and Conv() represents the convolution operation; by introducing global max pooling and global average pooling in parallel, the ability to perceive significant hotspot contours of power equipment in infrared images is enhanced.

[0063] Complementary mapping and feature aggregation: using the w1 to pair spatial features X SCalibration is performed using w2 on semantic feature X. C The calibration is performed, and the calibrated two feature streams are concatenated and aggregated to obtain the output feature X. FCM The calculation formula is as follows:

[0064] ;

[0065] In the formula, Concat() represents the channel concatenation operation.

[0066] Furthermore, all bottleneck units within the C3k2 module in the four stages from Stage 2 to Stage 5 of the backbone network are replaced with the AFCM module to achieve deep complementarity and alignment of spatial geometric features and semantic attribute features throughout the feature extraction process.

[0067] In a specific example, the implementation logic of the DCG-Fusion module is as follows:

[0068] Deep features x and shallow features y are mapped to the same dimension, and their difference feature term (xy) is calculated to explicitly extract abrupt changes in target edges in infrared images. The difference feature term is then modeled in a multi-scale space using a parallel dilated convolution sequence with multiple dilation rate branches. Pixel-level dynamic gating weights G are generated via an activation function, and their calculation formula is as follows:

[0069] ;

[0070] In the formula, DConv seq () represents a parallel dilated convolution sequence operation, used to adaptively adjust the feature receptive field to match power equipment of different scales; finally, the pixel-level dynamic gating weights G are used to perform complementary weighted fusion of deep and shallow features to obtain the output feature map Y, the calculation formula of which is:

[0071] .

[0072] Furthermore, the DCG-Fusion module is deployed at the intersection of deep and shallow features in the neck network, and dynamic content-guided fusion is used to solve the problems of large target scale span and noise interference in infrared images of power equipment.

[0073] In a specific example, the implementation logic of IPD-Loss is as follows:

[0074] Auxiliary box scaling: A scale factor r is introduced to synchronously scale the width and height of the predicted box and the ground truth box to construct an internal auxiliary box that focuses on the core region of the target.

[0075] Auxiliary point distance calculation: Calculate the squared Euclidean distance between corresponding corner points of the inner auxiliary boxes. The calculation formula is as follows:

[0076] ;

[0077] In the formula, and These represent the coordinates of the top-left corner (n=1) and bottom-right corner (n=2) of the predicted bounding box and the true bounding box, respectively.

[0078] Loss calculation: combining the intersection-union ratio (IoU) of the auxiliary bounding box. inner By mapping the point distance term, we obtain the regression loss function L. IPD :

[0079] ;

[0080] In the formula, IoU inner This represents the intersection-over-union ratio (IoU) between the scaled predicted bounding box and the ground truth bounding box. and , respectively represent the squared distance between corresponding corner points, w and h represent the original image size; the IPD-Loss enhancement network is used to improve the shape fitting ability of power equipment with different aspect ratios and multi-scale, and to suppress the influence of infrared image noise on positioning.

[0081] Step 3: Input the infrared image of the power equipment to be detected into the improved YOLO11 network with the best performance obtained from training, and output the power equipment detection results.

[0082] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for infrared image detection of power equipment based on an improved YOLO11 network, characterized in that, The improved YOLO11 network improves the backbone, neck network, and loss function of the original network. The improvements to the backbone include: constructing an improved feature calibration module (AFCM) within the backbone; the AFCM module extracts heterogeneous features through decoupled semantic and spatial branches and utilizes a dual-criteria weight extraction mechanism based on global average pooling and global max pooling to suppress infrared noise and enhance key equipment features; the improvements to the neck network include: constructing a dynamic content-guided fusion module (DCG-Fusion) to replace the original feature splicing structure; the DCG-Fusion module extracts mutation information by calculating the difference features between deep and shallow layers and uses parallel dilated convolution sequences to generate multi-scale gated weights for efficient perception of power targets of different sizes; and the improvements to the loss function include: constructing a kernel point distance loss (IPD-Loss); the IPD-Loss constructs internal auxiliary boxes using scale factors to focus on the core region of the target and utilizes the corner Euclidean distance mapping between the predicted and true auxiliary boxes to improve the shape fitting ability for irregularly shaped power equipment. The specific implementation of the infrared image detection method for power equipment includes the following steps: Step 1: Collect raw infrared images of power equipment, and after image enhancement and annotation, construct a power infrared detection dataset and perform image preprocessing to meet the requirements of the improved YOLO11 network input. Step 2: Input the power infrared detection dataset into the improved YOLO11 network for training. During training, the images in the training set are decoupled and extracted by the AFCM module, and multi-scale dynamic fusion is performed in the neck network by the DCG-Fusion module. During backpropagation, the regression loss is calculated using the IPD-Loss and the network parameters are updated to train the improved YOLO11 network with the best performance. Step 3: Input the infrared image of the power equipment to be detected into the improved YOLO11 network with the best performance obtained from training, and output the power equipment detection results.

2. The infrared image detection method for power equipment based on an improved YOLO11 network according to claim 1, characterized in that, Step 1 includes the following steps: Step 1.1, Power Infrared Data Acquisition: Use an infrared thermal imager to capture images of the power operation scene and obtain the original infrared image sequence containing the power equipment; Step 1.2, Infrared image quality enhancement: Adaptive histogram equalization or bilateral filtering algorithm is used to enhance the contrast and edge details of all infrared images in the original infrared image sequence; Step 1.3, Labeling and Sample Construction: Label the power equipment in the enhanced infrared image, generate a label file containing the target category and location coordinates, and construct a power infrared detection dataset; Step 1.4, Multi-scale mapping and normalization: The images and corresponding labels in the power infrared detection dataset are uniformly adjusted to a preset pixel size, and pixel value normalization is performed to make them meet the requirements of the improved YOLO11 network input.

3. The infrared image detection method for power equipment based on an improved YOLO11 network according to claim 1, characterized in that, The specific implementation logic of the AFCM module is as follows: Heterogeneous feature channel segmentation: The input feature map X is segmented into a first feature branch and a second feature branch in the channel dimension by a ratio α. By decoupling the category semantics and spatial location information in the initial stage, the decoupled first feature branch is used as the semantic branch and the decoupled second feature branch is used as the spatial branch, so as to improve the network's ability to retain low-level spatial information. Heterogeneous branch feature transformation: Standard convolution is performed on the semantic branches to obtain semantic features X containing rich category attributes. C The spatial branches are subjected to point-by-point convolution to obtain spatial features X that preserve the geometric topology. S ; A dual-criteria weight extraction mechanism based on global average pooling and global max pooling: based on the semantic feature X C With spatial features X S The channel weights w1 and spatial weights w2 used for complementary calibration are generated, and their calculation formulas are as follows: ； ； In the formula, σ() represents the Sigmoid activation function, MLP() represents the weight mapping operation of the multilayer perceptron, GAP() and GMP() represent global average pooling and global max pooling operations, respectively, and Conv() represents the convolution operation; by introducing global max pooling and global average pooling in parallel, the ability to perceive significant hotspot contours of power equipment in infrared images is enhanced. Complementary mapping and feature aggregation: using the w1 to pair spatial features X S Calibration is performed using w2 on semantic feature X. C The calibration is performed, and the calibrated two feature streams are concatenated and aggregated to obtain the output feature X. FCM The calculation formula is as follows: ； In the formula, Concat() represents the channel concatenation operation.

4. The infrared image detection method for power equipment based on an improved YOLO11 network according to claim 3, characterized in that, The bottleneck units within the C3k2 module in all four stages (Stage 2 to Stage 5) of the backbone network are replaced with the AFCM module to achieve deep complementarity and alignment of spatial geometric features and semantic attribute features throughout the feature extraction process.

5. The infrared image detection method for power equipment based on an improved YOLO11 network according to claim 1, characterized in that, The specific implementation logic of the DCG-Fusion module is as follows: Deep features x and shallow features y are mapped to the same dimension, and their difference feature term (xy) is calculated to explicitly extract abrupt changes in target edges in infrared images. The difference feature term is then modeled in a multi-scale space using a parallel dilated convolution sequence with multiple dilation rate branches. Pixel-level dynamic gating weights G are generated via an activation function, and their calculation formula is as follows: ； In the formula, DConv seq () represents a parallel dilated convolution sequence operation, used to adaptively adjust the feature receptive field to match power equipment of different scales; finally, the pixel-level dynamic gating weights G are used to perform complementary weighted fusion of deep and shallow features to obtain the output feature map Y, the calculation formula of which is: 。 6. The infrared image detection method for power equipment based on an improved YOLO11 network according to claim 5, characterized in that, The DCG-Fusion module is deployed at the intersection of deep and shallow features in the neck network. Through dynamic content-guided fusion, it solves the problems of large target scale span and noise interference in infrared images of power equipment.

7. The method for infrared image detection of power equipment based on an improved YOLO11 network according to claim 1, characterized in that, The specific implementation logic of IPD-Loss is as follows: Auxiliary box scaling: A scale factor r is introduced to synchronously scale the width and height of the predicted box and the ground truth box to construct an internal auxiliary box that focuses on the core region of the target. Auxiliary point distance calculation: Calculate the squared Euclidean distance between corresponding corner points of the inner auxiliary boxes. The calculation formula is as follows: ； In the formula, and These represent the coordinates of the top-left corner (n=1) and bottom-right corner (n=2) of the predicted bounding box and the actual bounding box, respectively. Loss calculation: combining the intersection-union ratio (IoU) of the auxiliary bounding box. inner By mapping the point distance term, we obtain the regression loss function L. IPD : ； In the formula, IoU inner This represents the intersection-over-union ratio (IoU) between the scaled predicted bounding box and the ground truth bounding box. and , respectively represent the squared distance between corresponding corner points, w and h represent the original image size; the IPD-Loss enhancement network is used to improve the shape fitting ability of power equipment with different aspect ratios and multi-scale, and to suppress the influence of infrared image noise on positioning.

Images