Power equipment infrared image detection method and device and medium

Abstract

The invention relates to the technical field of power equipment target detection, in particular to a power equipment infrared image detection method and device and a medium, and the method comprises the steps: constructing a DCLS-YOLOv8n model based on a YOLOv8n model, and replacing a C2F module in a backbone network of the YOLOv8n model with a residual module C2F-DCNv3 constructed by deformable convolution; adding a coordinate attention mechanism CA in the backbone network; designing a lightweight expanded feature pyramid LDFPN to replace a neck network of a YOLOv8n model, and adopting an SIoU frame regression loss function for training; and acquiring an infrared image of the power equipment, and identifying the infrared image of the power equipment by using the trained DCLS-YOLOv8n model. Compared with the prior art, the method has the advantages of high detection precision, low model parameter quantity, low calculation quantity and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of electric power equipment detection, and in particular to an infrared image detection method for electric power equipment. Background Art

[0002] Infrared imaging technology is widely used in the power industry because of its small impact on light and weather, as well as its advantages of non-contact and non-stop measurement. However, in the face of massive infrared image data, manual recognition and analysis have the disadvantages of low efficiency, high false detection rate, and extreme reliance on prior knowledge of engineers. Therefore, fast and accurate detection of power equipment in infrared images plays an important role in subsequent thermal defect screening and improving inspection efficiency.

[0003] At present, deep learning has made great progress in the fields of image segmentation and target detection. Many researchers have applied deep learning methods to power equipment recognition tasks. Compared with traditional detection methods, deep learning-based detection methods have the advantages of high detection efficiency, strong generalization ability, and no need for manually designed feature extractors, which can greatly improve the efficiency of equipment detection and fault diagnosis. Detection methods based on deep learning can be divided into two-stage detection algorithms and one-stage detection algorithms. Although the recognition accuracy of the one-stage algorithm is slightly lower than that of the two-stage algorithm, since it does not need to generate a large number of candidate boxes and can directly predict the location and category of the target, the detection speed is significantly improved, which is particularly suitable for real-time recognition and detection tasks of infrared images of power equipment.

[0004] In the field of one-stage model recognition of infrared images of power equipment, some literatures designed a global information aggregation module to improve the YOLOv4 model, guiding the network to pay more attention to the area where the target equipment is located. Other literatures use the MSRCR algorithm to enhance the infrared images of power equipment based on the characteristics of low contrast and blurred texture details of infrared images, and send them to the improved YOLOv4 model for detection. Although the average recognition accuracy of 8 types of power equipment is improved, additional image preprocessing operations are also added. There are also literatures that replace the standard convolution in the backbone network with phantom convolution on the basis of YOLOv5 to reduce the number of model parameters, add efficient attention and multi-head attention to improve the target attention and the detection effect of occluded targets, and introduce efficient intersection over union (Efficient Intersection over Union EIoU) to accelerate model convergence and improve recognition accuracy. Although the above studies have improved the recognition accuracy of power equipment in infrared images under complex environments to varying degrees, there are still some shortcomings. On the one hand, the actual power equipment has diverse shapes, and the detection effect of some equipment with high shape similarity is not ideal. On the other hand, the computational complexity of the model has room for optimization. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art, such as overly complex model calculation and poor detection effect on equipment with high shape similarity, and to provide an infrared image detection method, device and medium for power equipment based on the DCLS-YOLOv8n model.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] As a first aspect of the present invention, a method for detecting infrared images of power equipment is provided, wherein the steps are as follows: obtaining an infrared image of the power equipment, and detecting the infrared image of the power equipment using a trained DCLS-YOLOv8n model.

[0008] The DCLS-YOLOv8n model is based on the YOLOv8n model, and uses the residual module C2F-DCNv3 constructed by deformable convolution to replace the C2F module in the backbone network of the YOLOv8n model; adds a coordinate attention mechanism CA to the backbone network; and replaces the neck network of YOLOv8n with a lightweight dilated feature pyramid LDFPN.

[0009] As a preferred technical solution, the residual module C2F-DCNv3 constructed using deformable convolution replaces the C2F module in the YOLOv8n model backbone network as follows:

[0010] In the standard convolution bottleneck within C2F, DCNv3 is used to replace the second standard convolution to construct a bottleneck structure containing deformable convolution DCNv3.

[0011] As a preferred technical solution, the deformable convolution DCNv3 groups the input features, uses deep separable convolution to generate offsets and modulation coefficients, and each group has a separate offset and modulation coefficient; and then generates the modulation coefficient through the Softmax function.

[0012] As a preferred technical solution, for the p0 pixel point on the output feature y, the deformable convolution DCNv3 calculation formula is expressed as:

[0013]

[0014] Among them, y is the output feature map, G represents the number of groups, K represents the number of sampling points, and x g represents the input feature map after grouping, w gk represents the convolution kernel weight of the kth sampling point in the gth group, m gk represents the modulation coefficient of the kth sampling point in the gth group, p k represents the kth sampling position in the convolution kernel, Δp gk Indicates the offset of the kth sampling point in the gth group.

[0015] As a preferred technical solution, the coordinate attention mechanism CA is as follows:

[0016] By performing one-dimensional average pooling in both horizontal and vertical directions, attention vectors in two directions are obtained;

[0017] The horizontal and vertical attention vectors are concatenated and then reduced in dimension through convolution.

[0018] After normalization and nonlinear activation function, the spatial information is encoded;

[0019] The attention vectors in the horizontal and vertical directions are split, and the attention weights in the two directions are generated by the activation function after the convolution is used to increase the dimension.

[0020] By multiplying the attention weights with the input feature map, we get the attention-corrected output feature map.

[0021] As a preferred technical solution, the lightweight dilated feature pyramid LDFPN is specifically as follows:

[0022] Receive three scale feature maps of 20×20, 40×40 and 80×80 output from the backbone network;

[0023] The number of channels of each feature map is adjusted through convolution to compress and unify the number of channels;

[0024] LDFPN adopts a bottom-up feature aggregation and top-down feature transfer strategy. In the bottom-up process, the high-level semantic information in the low-resolution features is gradually transferred to the shallow features through the fusion module of the dilated residual structure. In the top-down process, the rich spatial information in the shallow features is gradually transferred to the low-resolution features through the fusion module of the dilated residual structure, thus realizing the fusion of semantic information and spatial information.

[0025] The multi-scale features after LDFPN fusion are passed to the detection head for final target detection, where 20×20 scale features are used to detect large targets, 40×40 scale features are used to detect medium targets, and 80×80 scale features are used to detect small targets.

[0026] As a preferred technical solution, the fusion module of the dilated residual structure is specifically as follows:

[0027] The number of channels of the feature map is compressed to c through the convolution kernel in the CBS module, and multi-scale features are initially fused;

[0028] The multi-scale features with the number of channels c are multi-scaled to generate three sets of feature maps with different receptive fields and stack them together in the channel dimension.

[0029] As an optimal technical solution, the DCLS-YOLOv8n model uses a bounding box regression loss function SIoU to replace the CIoU loss in the YOLOv8n model. The bounding box regression loss function SIoU includes angle loss Λ, distance loss Δ, shape loss Ω and intersection-over-union loss IoU.

[0030] As a second aspect of the present invention, there is provided an infrared image detection device for electric power equipment, comprising a memory, a processor, and a program stored in the memory, wherein the processor implements the infrared image detection method for electric power equipment as described above when executing the program.

[0031] As a third aspect of the present invention, there is provided a storage medium having a program stored thereon, wherein the program, when executed, implements the infrared image detection method for electric power equipment as described above.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1) Based on the original YOLOv8n, the present invention adds a coordinate attention mechanism, a deformable convolution residual module, a lightweight dilated feature pyramid LDFPN and a SIoU border regression loss function to improve the detection accuracy of power equipment in infrared images while reducing the number of parameters and the amount of calculation of the model.

[0034] 2) The present invention uses deformable convolution to construct the residual module C2F-DCNv3, replacing the C2F modules at three scales of 80×80, 40×40 and 20×20 in the YOLOv8n backbone network, enhancing the model's feature extraction capability for power equipment of various shapes, and at the same time reducing the computational complexity of the model to a certain extent.

[0035] 3) The present invention adds a coordinate attention mechanism after the C2F module at two scales of 80×80 and 40×40 in the YOLOv8n backbone network, which can suppress background features and improve the model's ability to pay attention to power equipment in infrared images.

[0036] 4) The present invention designs a lightweight dilated feature pyramid LDFPN to replace the neck network of the original YOLOv8n, thereby realizing the effective fusion of semantic information and spatial information, which can reduce the number of model parameters and calculation amount while ensuring detection accuracy, and enhance the model's detection effect on multi-scale targets.

[0037] 5) The present invention adopts the SIoU loss function to improve the positioning accuracy of the predicted box by optimizing the aspect ratio penalty term between the predicted box and the true box. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1It is a flow chart of the infrared image detection method of power equipment based on the DCLS-YOLOv8n model of the present invention;

[0039] Figure 2 A schematic diagram of a deformable convolution provided for one embodiment of the present invention;

[0040] Figure 3 A diagram of a standard convolution bottleneck structure and a deformable convolution bottleneck structure provided for an embodiment of the present invention; 3a) standard convolution bottleneck, 3b) DCNv3 bottleneck;

[0041] Figure 4 A schematic diagram of a coordinate attention mechanism provided for one embodiment of the present invention;

[0042] Figure 5 A structural diagram of a lightweight dilated feature pyramid LDFPN provided by an embodiment of the present invention;

[0043] Figure 6 A structural diagram of a DWR fusion module in a lightweight dilated feature pyramid LDFPN provided by an embodiment of the present invention;

[0044] Figure 7 A DCLS-YOLOv8n model structure diagram provided for one embodiment of the present invention;

[0045] Figure 8 A diagram of detection results provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0047] Example 1

[0048] like Figure 1 As shown, this embodiment provides a method for infrared image detection of electric power equipment based on the DCLS-YOLOv8n model, comprising the following steps:

[0049] S1. Build the DCLS-YOLOv8n model.

[0050] The DCLS-YOLOv8n model constructed by the present invention is based on the YOLOv8n model, uses the residual module C2F-DCNv3 constructed by deformable convolution to replace the C2F module in the backbone network of the YOLOv8n model, adds the coordinate attention mechanism CA to the backbone network, designs a lightweight dilated feature pyramid LDFPN to replace the neck network of the YOLOv8n model, and uses the SIoU border regression loss function for training

[0051] like Figure 2 As shown in the figure, in the deformable convolution residual module C2F-DCNv3, the deformable convolution DCNv3 uses deep separable convolution to generate offsets and modulation coefficients to reduce the number of parameters. Secondly, the input features are grouped, and each group has a separate offset and modulation coefficient, which can better learn the features of targets of different shapes. Finally, the Sigmoid function that generates the modulation coefficient is changed to Softmax, and the sum of the modulation coefficients is constrained to 1, making the model training process more stable. For the p0 pixel on the output feature y, the DCNv3 calculation formula is expressed as:

[0052]

[0053] Among them, y is the output feature map, G represents the number of groups, K represents the number of sampling points, and x g represents the input feature map after grouping, w gk represents the convolution kernel weight of the kth sampling point in the gth group, m gk represents the modulation coefficient of the kth sampling point in the gth group, p k represents the kth sampling position in the convolution kernel, Δp gk Indicates the offset of the kth sampling point in the gth group.

[0054] In order not to affect the model prediction speed too much, choose Figure 3 In the standard convolution bottleneck in C2F shown in (a), DCNv3 is used to replace the second standard convolution to construct a bottleneck structure containing DCNv3, as shown in Figure 3 (b) C2F-DCNv3 can enhance the model’s ability to extract adaptive features for devices of various shapes, while reducing the model’s computational complexity to a certain extent.

[0055] like Figure 4 As shown in the figure, the coordinate attention mechanism CA first obtains the attention vectors in two directions by performing one-dimensional average pooling in the horizontal and vertical directions. Then the two vectors are concatenated and reduced in dimension by 1×1 convolution, where r represents the dimensionality reduction multiple. Then the spatial information is encoded by normalization and ReLU nonlinear activation in sequence. Then the attention vectors in the two directions are split, and the dimensionality is increased by 1x1 convolution and then the attention weights in the two directions are generated by the Sigmoid function. Finally, the attention weights are multiplied by the input feature map to obtain the output feature map after attention correction.

[0056] The neck network of the original YOLOv8n is replaced with a lightweight dilated feature pyramid network (LDFPN) to further improve the multi-scale feature fusion capability and detection accuracy of the model and reduce the computational complexity of the model. The overall structure of LDFPN is as follows Figure 5 As shown in the figure, the three scale feature maps of 20×20, 40×40 and 80×80 output from the backbone network are received. The number of channels of each feature map is first adjusted through 1×1 convolution, and the number of channels is uniformly compressed to 64 to reduce the number of parameters and calculations of the model. On this basis, LDFPN adopts a bottom-up feature aggregation and top-down feature transfer strategy to achieve efficient fusion and refinement of multi-scale features. The bottom-up process gradually transfers high-level semantic information in low-resolution features to shallow features, while the top-down process gradually transfers rich spatial information in shallow features to low-resolution features, thereby achieving effective fusion of semantic information and spatial information and enhancing the model's detection effect on multi-scale targets.

[0057] The structure of the core module dilation-wise residual (DWR) in LDFPN is as follows Figure 6 As shown in the figure, the number of channels of the feature map is compressed to c through the CBS module, and the multi-scale features are initially fused. Then they are input into the dilated convolution with different expansion rates respectively, and three sets of feature maps with different receptive fields are generated and stacked together in the channel dimension. In this way, the features from different scales can be refined, and more extensive contextual information can be obtained through dilated convolution with different expansion rates, thereby improving the expression ability of multi-scale features. The multi-scale features fused by LDFPN are passed to the detection head for the final target detection, where the 20×20 scale features are used to detect large targets, the 40×40 scale features are used to detect medium targets, and the 80×80 scale features are used to detect small targets. By introducing LDFPN, ​​the model not only effectively improves the detection effect of multi-scale targets, but also reduces the overall computational complexity.

[0058] Figure 7 Schematic diagram of the overall structure of the DCLS-YOLOv8n model of the present invention.

[0059] The detection head of the basic YOLOv8 model contains two branches: classification and bounding box regression. The classification branch uses the BCE classification loss function to optimize the classification accuracy of the target category; the bounding box regression branch uses the CIoU bounding box regression loss function to optimize the position and shape of the predicted box so that it better fits the target of interest.

[0060] The DCLS-YOLOv8n model proposed in this invention uses the SIoU border regression loss function to replace the CIoU loss in the original model in order to further improve the positioning accuracy and robustness of the model during detection. The SIoU border regression loss function consists of angle loss (Λ), distance loss (Δ), shape loss (Ω) and intersection over union loss (IoU). The calculation formula for Λ is:

[0061]

[0062] Among them, β represents the angle between the line connecting the center point of the predicted box and the real box and the horizontal direction. During the training process, when β < π / 4, β will be minimized, and the predicted box will be close to the real box in the horizontal direction; when β ≥ π / 4, π / 2-β will be minimized, and the predicted box will continue to be close to the real box in the vertical direction. According to the above angle loss, the calculation formula of Δ is defined as:

[0063]

[0064]

[0065] Among them, b(b cx ,b cy )and Represents the center point coordinates of the predicted box and the real box, C w With C h Represent the width and height of the minimum bounding box respectively. The calculation formula of Ω is defined as:

[0066]

[0067]

[0068] Among them, w, h and w gt ,h gt Represents the width and height of the real box and the predicted box respectively. The calculation formula of SIoU is:

[0069]

[0070] S2. Obtain infrared images of power equipment, annotate and enhance the images, build a data set, and divide the data set into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0071] Specifically, 2687 infrared images of power equipment were sorted out, and 8 representative types of equipment were selected for annotation, including arrester, bushing, current limiter, fuse, insulator, knife, voltage transformer (PT) and current transformer (CT), where the English name is the label name of the corresponding target. In order to enhance the generalization ability of the model, the imgaug library is used in this embodiment to perform data enhancement on the original infrared image. The enhancement operations include random rotation of 90°, 180° or 270°, random scaling (scaling ratio between 0.7 and 0.95) and contrast adjustment (adjusted to 70% of the original contrast). In the process of image enhancement, each image is first rotated, and then scaling and contrast adjustment operations are performed with a probability of 50%, and each original image generates an enhanced new image. In this way, the original 2687 images were expanded to 5374 and divided into training set, validation set and test set in the ratio of 8:1:1.

[0072] S3. Configure the deep learning environment and set the DCLS-YOLOv8n model parameters.

[0073] In this embodiment, the operating system used in the experiment is Windows 10 64-bit, the processor is Intel(R) Core(TM) i5-13600KF, the graphics card is NVIDIAGeForce RTX 4070, the RAM size is 32G, the Python version used in the experiment is 3.9, the CUDA version is 11.7, and the deep learning framework is PyTorch2.0.1. During training, the input image will be uniformly adjusted to 640×640 pixels, and Mosaic data enhancement will be turned on. The model training cycle is set to 200 rounds, the batch size is set to 16, the number of multi-threads is 2, and mixed half-precision training is turned on. The model uses a stochastic gradient descent optimizer with an initial learning rate of 0.01 and a momentum of 0.937. The model uses a weight decay strategy to prevent overfitting, and the decay value is set to 5×10 -4 .

[0074] S4. Input the training set of infrared images of power equipment into the DCLS-YOLOv8n model for training. During the training process, the validation set is used for verification, and the trained model is tested with the test set.

[0075] This embodiment uses mean average precision (mAP), model parameters (Params) and giga floating point operations (GFLOPs) as evaluation indicators. The calculation method of mean average precision is:

[0076]

[0077]

[0078]

[0079]

[0080] Where R is the recall rate, P is the precision, TP is the number of correctly predicted positive samples, FN is the number of incorrectly detected positive samples, FP is the number of incorrectly predicted negative samples, AP is the area of ​​the precision and recall curve, and N is the number of detection categories.

[0081] S5. Use the trained DCLS-YOLOv8n model to detect infrared images of power equipment.

[0082] Figure 8 This is a detection result diagram obtained by using the method of the present invention in an embodiment. In this implementation, the trained DCLS-YOLOv8n model is evaluated, and it can be found that the mAP of the present invention reaches 94.92% on the self-built power equipment infrared data set, the model parameter volume is 1.7M, and the calculation amount is 6.1GFLOPs, which effectively improves the detection accuracy of the infrared image of the power equipment and reduces the calculation complexity of the detection model.

[0083] Example 2

[0084] This embodiment provides an infrared image detection device for electric power equipment based on a DCLS-YOLOv8n model, including a memory, a processor, and a program stored in the memory. When the processor executes the program, the method described in the above embodiment 1 is implemented.

[0085] In a preferred embodiment, the device comprises:

[0086] A model building module is used to build a DCLS-YOLOv8n model. The DCLS-YOLOv8n model backbone network includes a CBS module, a C2F module, an SPPF module, a residual module C2F-DCNv3 built with deformable convolution, and a coordinate attention mechanism CA. A lightweight dilated feature pyramid LDFPN is proposed to replace the neck network of the original model. The DCLS-YOLOv8n model adopts a SIoU border regression loss function.

[0087] The data set construction module is used to obtain infrared images of power equipment and perform annotation and image enhancement to construct a data set;

[0088] Environment and parameter configuration module, used to configure the deep learning environment and set DCLS-YOLOv8n model parameters;

[0089] Model training and testing module, used to input the infrared image dataset of power equipment into the DCLS-YOLOv8n model for training and testing;

[0090] The detection module is used to detect infrared images of power equipment using the trained DCLS-YOLOv8n model.

[0091] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0092] Example 3

[0093] This embodiment provides a storage medium on which a program is stored. When the program is executed, the method described in the above embodiment 1 is implemented.

[0094] Storage media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined in the present invention, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0095] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for infrared image detection of electric power equipment, characterized in that: The method steps are as follows: obtaining an infrared image of the power equipment, and detecting the infrared image of the power equipment using the trained DCLS-YOLOv8n model; The DCLS-YOLOv8n model is based on the YOLOv8n model, and uses the residual module C2F-DCNv3 constructed by deformable convolution to replace the C2F module in the backbone network of the YOLOv8n model; adds a coordinate attention mechanism CA to the backbone network; and designs a lightweight dilated feature pyramid LDFPN to replace the neck network of YOLOv8n.

2. The infrared image detection method for electric power equipment according to claim 1, characterized in that: The deformable convolution DCNv3 groups the input features and uses depth-separable convolution to generate offsets and modulation coefficients, with each group having a separate offset and modulation coefficient; and then generates the modulation coefficient through the Softmax function.

3. The infrared image detection method for electric power equipment according to claim 2, characterized in that: The residual module C2F-DCNv3 constructed using deformable convolution is specifically: in the standard convolution bottleneck in C2F, the second standard convolution is replaced by DCNv3 to construct a bottleneck structure containing deformable convolution DCNv3.

4. The infrared image detection method for electric power equipment according to claim 3 is characterized in that: For the p0 pixel on the output feature y, the deformable convolution DCNv3 calculation formula is expressed as: Among them, y is the output feature map, G represents the number of groups, K represents the number of sampling points, and x g represents the input feature map after grouping, w gk represents the convolution kernel weight of the kth sampling point in the gth group, m gk represents the modulation coefficient of the kth sampling point in the gth group, p k represents the kth sampling position in the convolution kernel, Δp gk Indicates the offset of the kth sampling point in the gth group.

5. The infrared image detection method for electric power equipment according to claim 1, characterized in that: The attention mechanism CA of the coordinates is as follows: The input feature map is subjected to one-dimensional average pooling in both horizontal and vertical directions to obtain the attention vectors in two directions; The horizontal and vertical attention vectors are concatenated and reduced in dimension through convolution, and then normalized and nonlinearly activated to encode spatial information. The attention vectors in the horizontal and vertical directions are split, and the attention weights in the two directions are generated by the activation function after the convolution is used to increase the dimension. By multiplying the attention weights with the input feature map, we get the attention-corrected output feature map.

6. The infrared image detection method for electric power equipment according to claim 1, characterized in that: The lightweight dilated feature pyramid LDFPN is as follows: Receive three scale feature maps of 20×20, 40×40 and 80×80 output from the backbone network; The number of channels of each feature map is adjusted through convolution to compress and unify the number of channels; A fusion module composed of an expanded residual structure is used to perform bottom-up feature aggregation and top-down feature transfer strategies; the bottom-up process gradually transfers high-level semantic information in low-resolution features to shallow features through the fusion module of the expanded residual structure; the top-down process gradually transfers rich spatial information in shallow features to low-resolution features through the fusion module of the expanded residual structure, thereby achieving fusion of semantic information and spatial information; The multi-scale features after LDFPN fusion are passed to the detection head for final target detection, where 20×20 scale features are used to detect large targets, 40×40 scale features are used to detect medium targets, and 80×80 scale features are used to detect small targets.

7. The infrared image detection method for electric power equipment according to claim 6, characterized in that: The fusion module of the dilated residual structure is specifically as follows: The number of channels of the feature map is compressed to c through the convolution kernel in the CBS module, and multi-scale features are initially fused; The multi-scale features with the number of channels c are used to generate three sets of feature maps with different receptive fields, and they are stacked together in the channel dimension to obtain the fusion results of features of different scales.

8. The infrared image detection method for electric power equipment according to claim 1, characterized in that: The DCLS-YOLOv8n model uses the bounding box regression loss function SIoU to replace the CIoU loss in the YOLOv8n model. The bounding box regression loss function SIoU includes angle loss Λ, distance loss Δ, shape loss Ω and intersection-over-union loss IoU.

9. An infrared image detection device for electric power equipment, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the infrared image detection method for electric power equipment as described in any one of claims 1-8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the infrared image detection method for electric power equipment as described in any one of claims 1-8 is implemented.

Images