Substation lead defect detection method and system based on monocular depth estimation

Abstract

The invention belongs to the technical field of visual defect detection, and particularly relates to a transformer substation lead defect detection method and system based on monocular depth estimation. The method comprises the following steps: collecting visible light and depth images of transformer substation lead equipment in normal and defect states, and constructing a data set after preprocessing; constructing a monocular depth estimation algorithm model for performing depth estimation on the visible light image to generate a corresponding depth image; respectively constructing defect detection models based on the visible light image and the depth image; and respectively obtaining defect detection results of the visible light image and the depth image, and combining defect identification results of the visible light image and the depth image. According to the method, the problem of relatively high false alarm of the visible light image caused by illumination, shielding and the like under a complex background is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of visual defect detection, and more specifically, relates to a method and system for detecting substation lead defects based on monocular depth estimation. Background Technology

[0002] Substations are common infrastructure for power transmission. Substation conductors are generally installed outdoors, and broken strands in substation conductors pose a major threat to the safe operation of the power system. Their dangers extend to multiple aspects, including power transmission, equipment operation, and personnel safety.

[0003] For example, Chinese patent document CN116091448A discloses a method for detecting broken strands in a conductor. Specifically, after acquiring an image, the method uses the YOLOv5 algorithm to identify the target conductor region. After color model conversion and grayscale processing, the conductor edges are extracted, and the conductor body is segmented. The method detects whether the conductor has broken strands by judging whether the area ratio of the segmented portion exceeds a set threshold range. Chinese patent document CN107886493A discloses a method for detecting strand defects in transmission line conductors. Specifically, it involves preprocessing the image to segment the conductor and background image, extracting features from the conductor image, and then identifying the strands.

[0004] The above methods enable the identification of broken or loose strands in conductors based on visible light images. However, in complex substation environments and harsh weather conditions, they are prone to false positives and false negatives due to image blurring and incomplete segmentation. Summary of the Invention

[0005] The present invention aims to overcome at least one of the defects of the prior art and provide a substation lead defect detection method based on monocular depth estimation, so as to solve the problem that false detection and missed detection are easily caused by image blurring and incomplete segmentation in the complex background and harsh weather conditions of substations.

[0006] The detailed technical solution of this invention is as follows: A substation lead defect detection method based on monocular depth estimation, the method comprising: S1. Collect visible light and corresponding depth images of the substation lead equipment in normal and defective states, and construct a training dataset after preprocessing. S2. Construct a monocular depth estimation algorithm model to generate a corresponding depth image from the visible light image to be detected, and train the model using the training dataset. S3. Construct a visible light image defect detection model and a depth image defect detection model respectively, perform lead wire defect detection on the visible light image and the depth image, and obtain the lead wire defect location region in the visible light image and the lead wire defect location region in the depth image. S4. Merge the lead defect location region of the acquired visible light image and the lead defect location region of the depth image, and determine whether the two lead defect location regions overlap. If yes, then the non-maximum suppression algorithm is used for post-processing, and highly overlapping detection boxes are merged based on the overlap threshold to return the optimal lead defect bounding box; if no, then the output results are directly merged to return the non-repeating lead defect region.

[0007] Furthermore, the monocular depth estimation algorithm model includes an encoder and a decoder, specifically: The encoder part uses ResNet as the base network and introduces dilated spatial pyramid pooling to capture multi-scale contextual information and extract multi-scale semantic and texture features of visible light images. It captures multi-scale contextual information by using dilated convolutions with different dilation rates in parallel, while maintaining the feature map resolution. The decoder uses the BRA attention mechanism to enhance the ability to extract key features. It decodes the features output by the encoder into a continuous depth map through Conv convolution and Sigmoid activation function. BRA addresses the problem of wasted computing and storage resources when processing large-scale data by introducing a dynamic, query-aware sparse attention mechanism.

[0008] Furthermore, the hollow space pyramid pooling specifically refers to: Multiple dilated convolutions with different dilation rates are applied in parallel to the deep feature maps output by the ResNet network. At the same time, the global average pooling result of the original deep feature maps is preserved and the number of channels is adjusted by 1×1 convolution. Then, the multi-scale convolutional features and global context features are concatenated by channel dimension and the number of channels is compressed by 1×1 convolution to capture feature information at different scales. The encoder finally outputs the multi-scale feature map after ASPP fusion, which serves as the input to the decoder (BRA).

[0009] Furthermore, the BRA attention mechanism is optimized, specifically including: The input ASPP fused multi-scale feature data is divided into: Several different regions, each region containing There are eigenvectors, where... , The feature data consists of height and width, which are then linearly mapped to obtain three inputs. , and ,right , Find the average value. , ,pass , Construct the adjacency matrix vector of regional correlation: In the formula, To construct the adjacency matrix vector, Represents the transpose of a matrix; Preserve the front of each area The index of the connections In the formula, For the index matrix, clustering , , , After aggregation Using attention operations: In the formula, Let be the attention function. Parameterization is achieved using depthwise separable convolutions. It is a normalized exponential function.

[0010] Furthermore, in S2, the monocular depth estimation algorithm model is constructed, and the loss function used is as follows: A hybrid loss mechanism combining Scale-Invariant Loss and Edge-Aware Loss is employed, with the weights of the monocular depth estimation algorithm model adjusted via error backpropagation. The loss function is as follows: : The Specifically: (1); In formula (1), To predict the depth map, This represents the true depth map, where N is the number of pixels and i is the pixel coordinate index in the vertical direction. To balance the hyperparameters; The Specifically: (2); In formula (2), For the horizontal Sobel operator, Let I be the Sobel operator in the vertical direction, and let I be the visible light image. represents the image gradient weights, where i is the pixel coordinate index in the vertical direction and j is the pixel coordinate index in the horizontal direction; To predict the pixel value at position (i,j) in the depth map Let be the pixel value of the true depth map at position (i,j). Let be the pixel value of the visible light image at position (i,j).

[0011] Furthermore, the construction of the visible light image defect detection model and the depth image defect detection model respectively includes: In the visible light image defect detection model, YOLOv7 is used as the base model, and a small target detection anchor is added to improve the small target detection capability of lead wire defects.

[0012] In the depth image defect detection model, YOLOv7 is used as the base model. First, the original depth map is preprocessed to extract its gradient map and surface normal map, which are then fused into a three-channel synthetic input to enhance the geometric feature representation. At the same time, an improved NMS strategy is adopted in the post-processing stage. The improved NMS strategy introduces the average depth difference between the target boxes as a constraint. If the average depth difference between the target boxes exceeds a threshold, it is not suppressed, thereby effectively improving the detection capability of overlapping or adjacent targets.

[0013] Furthermore, when the lead defect location region in the visible light image and the lead defect location region in the depth image overlap, the overlapping detection boxes are optimized and merged using a non-maximum suppression algorithm to output accurate defect localization results. When the detection results of the two modalities do not overlap, the two detection results are combined and the defect location result is output together.

[0014] In another aspect of the invention, a system for detecting substation lead defects based on monocular depth estimation is provided, the system comprising: Image acquisition module: Acquires visible light image data of the substation to be identified; Image conversion module: Equipped with a monocular depth estimation algorithm model, it converts the visible light image to be identified into the corresponding depth image; Image recognition module: Equipped with a constructed visible light image defect detection model and a depth image defect detection model, which respectively perform defect detection on the visible light and depth images to be identified, and respectively obtain the location region of the lead wire defect; Result processing module: This module merges the defective areas of the lead wires in the visible light image and depth image obtained by the image recognition module, and returns the detection results of the defective lead wire areas.

[0015] In another aspect of the invention, a computer-readable storage medium is also provided, which stores executable instructions that, when executed, cause the machine to perform a substation lead defect detection method based on monocular depth estimation as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: A method and system for detecting lead defects in a substation based on monocular depth estimation provided by the present invention generate a corresponding depth image from a visible light image, and respectively construct lead defect detection algorithm models based on visible light and depth images. By integrating the advantages of rich details in visible light images and little influence of depth images by illumination, the false detection rate caused by factors such as illumination changes and occlusion is reduced, and the accuracy and robustness of lead defect detection in complex scenes of substations are effectively improved. At the same time, a dedicated Anchor mechanism for small target detection is introduced into the improved defect detection algorithm model, which enhances the sensitivity of the model to small-sized defects and further improves the recognition accuracy of lead defect detection in substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flowchart of the defect detection method based on monocular depth estimation according to the present invention; Figure 2 It is a schematic diagram of the monocular depth estimation algorithm model in Embodiment 1 of the present invention; Figure 3 It is a schematic structural diagram of the BRA attention module in Embodiment 1 of the present invention; Figure 4 It is a flowchart of the lead defect detection method in Embodiment 1 of the present invention; Figure 5 It is a schematic diagram of the system structure in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The present invention will be further described below with reference to the drawings and embodiments.

[0019] It should be noted that the following detailed descriptions are all exemplary and are intended to provide further descriptions of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0020] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0021] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0022] Embodiment 1 Refer Figure 1 , this embodiment provides a method for detecting lead defects in a substation based on monocular depth estimation, and the method includes: S1. Collect visible light and depth images of the substation lead wire equipment under normal and defective conditions, and construct a training dataset after preprocessing. Defective conditions include broken strands and loose strands. Preferably, visible light and depth images of the substation lead equipment in normal and defective states are acquired; the preprocessing includes: registering and organizing the acquired visible light images and corresponding depth images to ensure data alignment, thereby constructing a dedicated dataset for training the monocular depth estimation algorithm, which is used for training and validating the monocular depth estimation network. Meanwhile, annotation tools were used to annotate visible light and depth images to detect lead defects, creating a lead defect target detection dataset, which was used for training and validation of lead defect target detection networks for visible light and depth images, respectively.

[0023] S2. Construct a monocular depth estimation algorithm model to generate a corresponding depth image from the visible light image to be detected, and train the model using the training dataset. Preferably, the monocular depth estimation algorithm model is constructed as follows: Figure 2 As shown, it specifically includes: The encoder part uses ResNet as its base network and introduces Atrous Spatial Pyramid Pooling (ASPP) to capture multi-scale contextual information and extract multi-scale semantic and textural features from visible light images. Multi-scale contextual information is explicitly acquired at multiple scales, enabling the model to effectively handle changes in object size. Diffuse convolutions are used to compute dense feature maps, which, compared to traditional upsampling / deconvolution methods alone, requires less memory or computation. Multi-scale contextual information is captured by using dilated convolutions with different dilation rates in parallel, while maintaining feature map resolution.

[0024] The decoder uses the BRA (Bi-Level Routing Attenton) attention mechanism to enhance the ability to extract key features. It decodes the features output by the encoder into a continuous depth map through Conv convolution and Sigmoid activation function. Preferably, the Atrous Spatial Pyramid Pooling (ASPP) specifically includes: Multiple dilated convolutions with different dilation rates (e.g., 1, 6, 12, 18) are applied in parallel to the deep feature maps output by the ResNet network. At the same time, the global average pooling result of the original deep feature maps is preserved and the number of channels is adjusted by 1×1 convolution. Then, the multi-scale convolutional features and global context features are concatenated by channel dimension, and the number of channels is compressed by 1×1 convolution to capture feature information at different scales. The encoder finally outputs the multi-scale feature map after ASPP fusion, which serves as the input to the decoder (BRA).

[0025] like Figure 3 As shown, the attention mechanism described is Bi-Level Routing Attenton (BRA) dynamic sparse attention, which is optimized through the following methods: The input ASPP fused multi-scale feature data is divided into: Several different regions, each region containing There are eigenvectors, where... , The feature data consists of height and width, which are then linearly mapped to obtain three inputs. , and ,right , Find the average value. , That is, the Mean function, through , Construct the adjacency matrix vector of regional correlation: In the formula, To construct the adjacency matrix vector, Represents matrix transpose. Figure 3 The Matmul function is used to calculate the matrix product of two arrays; Preserve the front of each area The index of the connections In the formula, For the index matrix, clustering , , , After aggregation Using attention operations: In the formula, Let be the attention function. Parameterization is achieved using depthwise separable convolutions. It is a normalized exponential function.

[0026] The feature map output from BRA is further refined through convolutional layers, and the feature mapping is normalized by the Sigmoig activation function, finally outputting a single-channel depth map.

[0027] The Sigmoig activation function is specifically as follows: (3); In formula (3), x is the output feature of the convolutional layer, and e is the base of the natural logarithm.

[0028] Preferably, in step S2, the loss function used to construct the monocular depth estimation algorithm model is as follows: A hybrid loss mechanism combining Scale-Invariant Loss and Edge-Aware Loss is employed, with the weights of the monocular depth estimation algorithm model adjusted via error backpropagation. The loss function is as follows: : The Specifically: (1); In formula (1), To predict the depth map, This represents the true depth map, where N is the number of pixels and i is the pixel coordinate index in the vertical direction. To balance the hyperparameters, they are usually set to 0.5.

[0029] The Specifically: (2); In formula (2), For the horizontal Sobel operator, Let I be the Sobel operator in the vertical direction, and let I be the visible light image. represents the image gradient weights, where i is the pixel coordinate index in the vertical direction and j is the pixel coordinate index in the horizontal direction; To predict the pixel value at position (i,j) in the depth map Let be the pixel value of the true depth map at position (i,j). Let be the pixel value of the visible light image at position (i,j).

[0030] S3. Construct a visible light image defect detection model and a depth image defect detection model, and perform lead wire defect detection.

[0031] Specifically, a visible light image defect detection model and a depth image defect detection model are constructed respectively. Lead defects are detected in both visible light and depth images to obtain the lead defect location regions in the visible light image and the lead defect location regions in the depth image. In this embodiment, YOLOv7 is used as the base model to construct a visible light image defect detection model and a depth image defect detection model, respectively.

[0032] In the visible light image defect detection model, YOLOv7 is used as the base model, and a small target detection anchor is added to improve the small target detection capability of lead wire defects.

[0033] In the depth image defect detection model, YOLOv7 is used as the base model. First, the original depth map is preprocessed to extract its gradient map and surface normal map, which are then fused into a three-channel synthetic input to enhance the geometric feature representation. At the same time, an improved NMS strategy is adopted in the post-processing stage. The improved NMS strategy introduces the average depth difference between the target boxes as a constraint. If the average depth difference between the target boxes exceeds a threshold, it is not suppressed, thereby effectively improving the detection capability of overlapping or adjacent targets.

[0034] Specifically, such as Figure 4 As shown, the visible light image to be identified is input into the monocular depth estimation algorithm model in step S2 to generate the corresponding depth image. The visible light image to be identified is input into the lead defect detection model based on the visible light image, and the corresponding depth image is input into the lead defect detection model based on the depth image. The lead defect location regions of the visible light image and the lead defect location regions of the depth image are obtained simultaneously.

[0035] S4. Determine whether the defect locations of the two leads overlap and output the detection results.

[0036] Specifically, the lead defect location regions in the acquired visible light image and the lead defect location regions in the depth image are merged to determine whether the identified lead defect location regions overlap. If yes, then the non-maximum suppression algorithm is applied for post-processing, and highly overlapping detection boxes are merged based on the overlap threshold to return the optimal lead defect bounding box. If not, the output results will be directly merged, and the non-repeating lead defect areas will be returned.

[0037] Preferably, in this embodiment, such as Figure 4 As shown, when the lead defect location region in the visible light image and the lead defect location region in the depth image overlap, the non-maximum suppression (NMS) algorithm is used to optimize and merge the overlapping detection boxes, and output accurate defect localization results. When the detection results of the two modes do not overlap, the two detection results are directly merged, and the non-repeating lead defect area is returned, that is, all defect location results are output.

[0038] In this embodiment, the detection rate and false detection rate were tested using a visible light image-based method for identifying broken or loose strands in wires, a depth image-based method for identifying broken or loose strands in wires, and the method based on the present invention. The specific experimental data results are shown in Table 1 below: Table 1. Experimental results of detection rate and false detection rate

[0039] Therefore, the present invention achieves a detection rate of 89.65%, which is 6.4% and 2.33% higher than the detection methods based on visible light images and depth images, respectively. The false detection rate is only 11.73%, which is 4.65% and 0.84% ​​lower than the detection methods based on visible light images and depth images, respectively.

[0040] Example 2 This embodiment provides a system for implementing a substation lead defect detection method based on monocular depth estimation, such as... Figure 5 As shown, the system includes: Image acquisition module: Acquires visible light image data of the substation to be identified; Image conversion module: Equipped with a monocular depth estimation algorithm model, it converts the visible light image to be identified into the corresponding depth image; Image recognition module: Equipped with a constructed visible light image defect detection model and a depth image defect detection model, which respectively perform defect detection on the visible light and depth images to be identified, and respectively obtain the location region of the lead wire defect; Result processing module: This module merges the defective areas of the lead wires in the visible light image and depth image obtained by the image recognition module, and returns the detection result of the only defective lead wire area.

[0041] Example 3 This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed, cause the machine to perform the substation lead defect detection method and system based on monocular depth estimation as described above.

[0042] Specifically, a system or apparatus equipped with a readable storage medium may be provided, on which software program code implementing the functions of any of the embodiments described above is stored, and the computer or processor of the system or apparatus can read and execute the instructions stored in the readable storage medium.

[0043] In this case, the program code itself, which can be read from the readable medium, can perform the functions of any of the above embodiments, and therefore the computer-readable code and the readable storage medium storing the computer-readable code constitute a part of this specification.

[0044] Examples of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer or the cloud via a communication network.

[0045] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0046] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0047] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0049] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A method for detecting substation lead defects based on monocular depth estimation, characterized in that, The method includes: S1. Collect visible light and corresponding depth images of the substation lead equipment in normal and defective states, and construct a training dataset after preprocessing. S2. Construct a monocular depth estimation algorithm model to generate a corresponding depth image from the visible light image to be detected, and train the model using the training dataset. S3. Construct a visible light image defect detection model and a depth image defect detection model respectively, perform lead wire defect detection on the visible light image and the depth image, and obtain the lead wire defect location region in the visible light image and the lead wire defect location region in the depth image. S4. Merge the lead defect location region of the acquired visible light image and the lead defect location region of the depth image, and determine whether the two lead defect location regions overlap. If so, the non-maximum suppression algorithm is used for post-processing, and highly overlapping detection boxes are merged based on the overlap threshold to return the optimal lead defect bounding box. If not, the output results will be directly merged, returning the non-repeating lead defect areas.

2. The substation lead defect detection method based on monocular depth estimation according to claim 1, characterized in that, The monocular depth estimation algorithm model includes an encoder and a decoder, specifically: The encoder part uses ResNet as the base network and introduces dilated spatial pyramid pooling to capture multi-scale contextual information and extract multi-scale semantic and texture features of visible light images. It captures multi-scale contextual information by using dilated convolutions with different dilation rates in parallel, while maintaining the feature map resolution. The decoder uses the BRA attention mechanism to enhance the ability to extract key features, and decodes the features output by the encoder into a continuous depth map through Conv convolution and Sigmoid activation function.

3. The substation lead defect detection method based on monocular depth estimation according to claim 2, characterized in that, The hollow space pyramid pooling specifically refers to: Multiple dilated convolutions with different dilation rates are applied in parallel to the deep feature maps output by the ResNet network. At the same time, the global average pooling result of the original deep feature maps is preserved and the number of channels is adjusted by 1×1 convolution. Then, the multi-scale convolutional features and global context features are concatenated by channel dimension and the number of channels is compressed by 1×1 convolution to capture feature information at different scales. The encoder finally outputs the multi-scale feature map after ASPP fusion as the input of the decoder.

4. The substation lead defect detection method based on monocular depth estimation according to claim 2, characterized in that, Optimize the BRA attention mechanism, specifically including: The input ASPP fused multi-scale feature data is divided into: Several different regions, each region containing There are eigenvectors, where... , The feature data consists of height and width, which are then linearly mapped to obtain three inputs. , and ,right , Find the average value. , ,pass , Construct the adjacency matrix vector of regional correlation: In the formula, To construct the adjacency matrix vector, Represents the transpose of a matrix; Preserve the front of each area The index of the connections In the formula, For the index matrix, clustering , , , After aggregation Using attention operations: In the formula, Let be the attention function. Parameterization is achieved using depthwise separable convolutions. It is a normalized exponential function.

5. The substation lead defect detection method based on monocular depth estimation according to claim 4, characterized in that, The monocular depth estimation algorithm model constructed in S2 uses the following loss function: A hybrid loss mechanism combining Scale-Invariant Loss and Edge-Aware Loss is employed, with the weights of the monocular depth estimation algorithm model adjusted via error backpropagation. The loss function is as follows: ; The Specifically: （1）； In formula (1), To predict the depth map, This represents the true depth map, where N is the number of pixels and i is the pixel coordinate index in the vertical direction. To balance the hyperparameters; The Specifically: （2）； In formula (2), For the horizontal Sobel operator, Let I be the Sobel operator in the vertical direction, and let I be the visible light image. represents the image gradient weights, where i is the pixel coordinate index in the vertical direction and j is the pixel coordinate index in the horizontal direction; To predict the pixel value at position (i,j) in the depth map Let be the pixel value of the true depth map at position (i,j). Let be the pixel value of the visible light image at position (i,j).

6. A substation lead defect detection method based on monocular depth estimation according to claim 4 or 5, characterized in that, The construction of visible light image defect detection models and depth image defect detection models respectively includes: In the visible light image defect detection model, YOLOv7 is used as the base model, and small target detection anchors are added to improve the small target detection capability of lead wire defects, thus constructing a visible light image lead wire defect target detection model. In the depth image defect detection model, YOLOv7 is used as the base model. First, the original depth map is preprocessed to extract its gradient map and surface normal map, which are then fused into a three-channel synthetic input to enhance the geometric feature representation. At the same time, an improved NMS strategy is adopted in the post-processing stage. The improved NMS strategy introduces the average depth difference between the target boxes as a constraint. If the average depth difference between the target boxes exceeds a threshold, it is not suppressed.

7. A system for implementing a substation lead defect detection method based on monocular depth estimation, characterized in that, The system includes: Image acquisition module: Acquires visible light image data of the substation to be identified; Image conversion module: Equipped with a monocular depth estimation algorithm model, it converts the visible light image to be identified into the corresponding depth image; Image recognition module: Equipped with a constructed visible light image defect detection model and a depth image defect detection model, which respectively perform defect detection on the visible light and depth images to be identified, and respectively obtain the location region of the lead wire defect; Result processing module: This module merges the defective areas of the lead wires in the visible light image and depth image obtained by the image recognition module, and returns the detection results of the defective lead wire areas.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images