Defect detection method and device, system and storage medium for overhead power communication optical cable
By improving the backbone network and feature fusion network, the problem of low detection accuracy of defects in overhead power communication optical cables in the existing technology has been solved, and efficient detection of defects in small target optical cables has been achieved, thus improving the quality of inspection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- INFORMATION & COMM CO OF STATE GRID JILIN ELECTRIC POWER CO LTD
- Filing Date
- 2025-11-14
- Publication Date
- 2026-07-03
Smart Images
Figure CN121504865B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of overhead power communication optical cable technology, specifically relating to a method, device, system, and storage medium for detecting defects in overhead power communication optical cables. Background Technology
[0002] Aerial power communication optical cables bear core functions such as transmitting power grid dispatch commands and transmitting equipment status monitoring data. If a cable is interrupted due to a defect, it can lead to power grid dispatch failure and even widespread power outages. If defects in these cables are not detected in time, they may worsen under severe weather conditions (strong winds, lightning strikes), increasing repair costs and potentially causing cascading economic losses such as business shutdowns and disruptions to public services. Therefore, regular inspections of aerial power communication optical cables are necessary to promptly identify and address defects, ensuring the reliability of power communication.
[0003] UAV-based inspection of overhead power and communication fiber optic cables has become an important method of routine inspection. UAVs, using their lightweight computing devices, can identify inspection images in real time, effectively improving inspection efficiency. However, defects in overhead power and communication fiber optic cables are relatively small and their features are not obvious enough. Existing defect detection methods struggle to extract sufficient features, resulting in low defect detection accuracy and affecting the quality of overhead power and communication fiber optic cable inspections. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention provides a method, device, system, and storage medium for detecting defects in overhead power communication optical cables.
[0005] To achieve the above objectives, the present invention provides the following solution:
[0006] A method for detecting defects in overhead power communication optical cables includes:
[0007] Step 1: Input the inspection image of the overhead power communication optical cable into the improved backbone network to extract the features of the communication optical cable at different scales; the improved backbone network consists of convolutional layer 1, convolutional layer 2, C1 module, convolutional layer 3, C2 module, convolutional layer 4, A2C2f module 1, convolutional layer 5, A2C2f module 2, depth feature extraction module and RFA module in sequence;
[0008] Step 2: Output the feature information of the four different scales output in Step 1 to the improved feature fusion network; the improved feature fusion network includes a first-layer fusion network and a second-layer fusion network. The first-layer fusion network consists of a CARAFEE module, a stitching operation, an A2C2f module 3, a CARAFEE module 2, a stitching operation, an A2C2f module 4, a CARAFEE module 3, and a stitching module in sequence; the second-layer fusion network consists of an A2C2f module 5, a convolutional layer 6, a convolutional layer 7, an A2C2f module 6, a convolutional layer 8, a stitching operation, an A2C2f module 7, a convolutional layer 9, a stitching operation, and a C3 module in sequence.
[0009] Step 3: Input the four fused features output from Step 2 into the detection network for detection, thereby realizing the detection of defects in power communication optical cables.
[0010] Preferably, the C1 and C2 modules in the improved backbone network have the same network structure. Their input feature maps are processed by a regular convolution with a kernel of 1 to adjust the number of channels, resulting in feature map F1. Feature map F1 is then split to obtain feature maps F2 and F3. Feature map F2 is then processed sequentially by a regular convolution with a kernel of 1, normalization, SiLU activation, a depthwise separable convolution with a kernel of 3, normalization, SiLU activation, a regular convolution with a kernel of 1, normalization, and SiLU activation to obtain feature map F4. Feature map F4 is then processed by... After passing through a regular convolution with a kernel of 1 and the SiLU activation function, the channel weights in the F4 feature map are obtained. These channel weights are multiplied by feature map F4 to obtain feature map F5. Feature map F5 is then processed by a channel adjustment module consisting of a regular convolution with a kernel of 1, normalization, the SiLU activation function, and a regular convolution with a kernel of 1, and then element-wise added to feature map F4 to obtain feature map F6. Feature map F6 is concatenated with feature maps F3 and F1 to obtain feature map F7. Finally, feature map F7 is processed by a regular convolution with a kernel of 1 to obtain the output feature map of module C1 or module C2.
[0011] Preferably, the deep feature extraction module consists of five branches. The first branch consists of a regular convolution with a kernel of 1 and a SiLU activation function. The second branch consists of a regular convolution with a kernel of 1, a SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 3. The third branch consists of a pinwheel convolution with a kernel of 3, a SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 3. The fourth branch consists of a pinwheel convolution with a kernel of 5, a SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 5. The fifth branch is a short connection used to store the input features. The output features from the first to the fourth branch are fused through a concatenation operation. The fused feature map is then processed by a regular convolution and a SiLU activation function and added element-wise with the input features of the module to obtain the final output feature map.
[0012] Preferably, the RFA module consists of two branches. The first branch consists of average pooling, a sigmoid function, and a CARAFFE module in sequence, and the output feature map of the first branch is E1. The second branch consists of a pinwheel convolution with a kernel of 3, and its output feature map is E2. Feature map E1 is multiplied element-wise with feature map E2, and the result is added element-wise to feature map E1 to obtain feature map E3. Feature map E3 is adjusted by regular convolution to obtain the output feature map of the RFA module.
[0013] Preferably, in step 2, the output feature map of module C3 is subjected to a regular convolution with a kernel of 1 to obtain feature map D1; feature map D1 is split into feature map D2 and feature map D3 using a splitting operation, with a channel ratio of 1:1 between feature map D2 and feature map D3; feature map D2 is subjected to the sixth branch to obtain feature map D4, and feature map D2 is subjected to the seventh branch to obtain feature map D5; feature map D4, feature map D5, and feature map D6 are concatenated to obtain feature map D7; feature map D7 is subjected to a regular convolution to adjust the channel dimension, and the output feature map of module C3 is output; the sixth branch consists of two residual modules with the same structure, and the main module of each residual module is composed of a regular convolution with a kernel of 1, normalization, SiLU activation function, a windmill-shaped convolution with a kernel of 3, normalization, SiLU activation function, a regular convolution with a kernel of 1, normalization, and SiLU activation function in sequence.
[0014] The present invention also provides a defect detection device for overhead power communication optical cables, comprising:
[0015] The first processing unit is used to input the inspection image of the overhead power communication optical cable into the improved backbone network and extract the features of the communication optical cable at different scales. The improved backbone network consists of convolutional layer 1, convolutional layer 2, C1 module, convolutional layer 3, C2 module, convolutional layer 4, A2C2f module 1, convolutional layer 5, A2C2f module 2, a depth feature extraction module, and an RFA module.
[0016] The second processing unit is used to output the feature information of four different scales from the first processing unit to the improved feature fusion network. The improved feature fusion network includes a first-layer fusion network and a second-layer fusion network. The first-layer fusion network consists of a CARAFEE module, a splicing operation, an A2C2f module 3, a CARAFEE module 2, a splicing operation, an A2C2f module 4, a CARAFEE module 3, and a splicing module in sequence. The second-layer fusion network consists of an A2C2f module 5, a convolutional layer 6, a convolutional layer 7, an A2C2f module 6, a convolutional layer 8, a splicing operation, an A2C2f module 7, a convolutional layer 9, a splicing operation, and a C3 module in sequence.
[0017] The third processing unit is used to input the four fused features output by the second processing unit into the detection network for detection, thereby realizing the detection of defects in power communication optical cables.
[0018] The present invention also provides an overhead power communication optical cable defect detection system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes an overhead power communication optical cable defect detection method when executed by the processor.
[0019] The present invention also provides a storage medium storing a computer program, which executes a method for detecting defects in overhead power communication optical cables when running.
[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0021] 1. Introduce modules C1 and C2 to better extract defect features of communication optical cables by combining global context information;
[0022] 2. A deep feature extraction module is introduced, which can extract deeper features of communication optical cables more comprehensively from multiple scales;
[0023] 3. Introducing the RFA module can enhance the local features of defects in communication optical cables and improve the location and boundary quality of defect detection.
[0024] 4. The introduction of the C3 module can better extract small target features from the fused low-resolution features, thereby improving the feature extraction capability for small targets. Attached Figure Description
[0025] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0026] Figure 1 This is a flowchart of the defect detection method for overhead power communication optical cables according to an embodiment of the present invention;
[0027] Figure 2 The diagram shows the network structure for defect detection of overhead transmission lines based on the improved YOLOv12s.
[0028] Figure 3 A structural diagram of the C1 or C2 module in the improved backbone network;
[0029] Figure 4 Network structure diagram of the deep feature extraction module in the improved backbone network;
[0030] Figure 5 Network structure diagram of the RFA module in the improved backbone network;
[0031] Figure 6 Network structure diagram of C3 module in the improved feature fusion network;
[0032] Figure 7 This is for detecting the network structure diagram of the graph network head. Detailed Implementation
[0033] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0035] Example 1
[0036] like Figures 1 to 6 As shown, the present invention provides a method for detecting defects in overhead power communication optical cables, comprising:
[0037] Step 1: Input the inspection image of the overhead power communication optical cable into the improved backbone network to extract features of the communication optical cable at different scales. The improved backbone network consists of convolutional layer 1, convolutional layer 2, C1 module, convolutional layer 3, C2 module, convolutional layer 4, A2C2f module 1, convolutional layer 5, A2C2f module 2, a depth feature extraction module, and an RFA module. The C1, C2, A2C2f, and RFA modules output feature information at four different scales. Convolutional layer 1 and convolutional layer 2 perform continuous downsampling to generate feature maps of the edges and shallow texture features of communication optical cable defects. This feature map is then fed into the C1 module. A downsampling process is performed in the C1 module. Downsampling and multi-branch feature extraction and fusion not only preserve the detailed texture features of the original input feature map but also superimpose local semantics. Meanwhile, skip connections ensure that defect features of small targets are not easily lost. The lateral output of module C1 also provides high-resolution edge and texture information for subsequent feature fusion networks and the detection head for small target defects. The output feature map of module C1 is fed into convolutional layer 3 for downsampling, generating a feature map that preserves edge details while suppressing background noise. This feature map is then input into module C2, which serves the same purpose as module C1, producing a more expressive feature map. The lateral output of module C2 helps in feature extraction. The fusion network retains details and supplements strong semantic information, improving the separability of localization and classification. The gradient channels of the residuals help the overall network train and converge more stably. The output feature map of module C2 is downsampled by convolutional layer 4 to obtain the mid-level semantic feature map. The feature map output from convolutional layer 4 is then processed by module A2C2f1, a multi-head region attention module. This module divides the input feature map according to regions, uniformly dividing it into four parts. Self-attention is performed on each of these four parts. The module generates queries / keys / values through QKV projection and calculates multi-head self-attention on the four divided feature maps. The attention is limited to local regions, thus strengthening the detection of defects in communication optical cables. The module improves the relationship between the defect and the background, reducing computational complexity. Simultaneously, the lateral output of this module is a top-down fusion of the fusion network, providing contextual enhancement while maintaining detailed representation, thus improving the recall and localization stability of small / medium-scale optical cable defects. The output feature map of module 1 is further downsampled by convolutional layer 5 to obtain a deep feature map. This deep feature map is then input into module 2, where self-attention is calculated on the divided regions. The final output feature map suppresses background texture noise, emphasizes the relationship between the defect and the surrounding background, and helps in identifying medium-to-large-scale optical cable defects.The feature map output from module A2C2f 2 is used as input to the deep feature extraction module. This module expands the receptive field without downsampling and emphasizes longitudinal, lateral, and diagonal structures through windmill convolution. It preserves original information and stabilizes gradients through residual summation. Its output feature map is characterized by its focus on details and the global picture, and its greater sensitivity to thin, long optical cables and broken strands. Finally, the output of the deep feature extraction module is used as input to the RFA module. In the RFA module, an adaptive spatial weight map is generated from the input feature map and multiplied with the output features of the windmill convolution in the RFA module. The resulting feature map further emphasizes defective areas, suppresses redundant features, and improves the quality of the feature map.
[0038] Step 2: Output the feature information of the four different scales output in Step 1 to the improved feature fusion network; the improved feature fusion network includes a first-layer fusion network and a second-layer fusion network. The first-layer fusion network consists of a CARAFEE module, a stitching operation, an A2C2f module 3, a CARAFEE module 2, a stitching operation, an A2C2f module 4, a CARAFEE module 3, and a stitching module in sequence; the second-layer fusion network consists of an A2C2f module 5, a convolutional layer 6, a convolutional layer 7, an A2C2f module 6, a convolutional layer 8, a stitching operation, an A2C2f module 7, a convolutional layer 9, a stitching operation, and a C3 module in sequence.
[0039] Step 3: Input the four fused features output from Step 2 into the detection network for detection, thereby realizing the detection of defects in power communication optical cables. The detection network consists of four detection heads, each with the same structure as the original detection head in YOLOv12s. The structure diagram of this detection head is shown below. Figure 7 As shown, the detection head has a bounding box regression branch (BBox branch) and a category classification branch (Cls branch). In the bounding box regression branch, the feature map fused from the feature pyramid passes through two 3*3 ordinary convolutional layers, which consist of ordinary convolution, normalization, and SiLU activation functions, followed by a 1*1 ordinary convolution to calculate the bounding box loss. The classification branch structure is similar to the regression branch, except that the output channel is the number of categories C=nc, and the category confidence corresponding to each grid position is output through the Sigmoid activation function.
[0040] In one embodiment of the present invention, convolutional layers 1 to 5 in the improved backbone network in step 1 are each composed of a pinwheel convolution, normalization, and SiLU activation function in sequence; the C1 and C2 modules in the improved backbone network have the same network structure, and their input feature maps are processed by a regular convolution with a kernel of 1 to adjust the number of channels of the input feature map, resulting in feature map F1; after splitting, feature map F1 is obtained as feature map F2 and feature map F3; feature map F2 is processed by a regular convolution with a kernel of 1, normalization, SiLU activation function, depthwise separable convolution with a kernel of 3, normalization, SiLU activation function, regular convolution with a kernel of 1, normalization, and SiLU activation function in sequence, resulting in feature map F4; feature map F4 is processed by a convolution kernel... After a regular convolution with a kernel of 1 and the SiLU activation function, the channel weights in the F4 feature map are obtained. These channel weights are multiplied by feature map F4 to obtain feature map F5. Feature map F5 is then processed by a channel adjustment module consisting of a regular convolution with a kernel of 1, normalization, the SiLU activation function, and a regular convolution with a kernel of 1. Finally, feature map F6 is added element-wise to feature map F4 to obtain feature map F6. Feature map F6 is concatenated with feature maps F3 and F1 to obtain feature map F7. Finally, feature map F7 is processed by a regular convolution with a kernel of 1 to obtain the output feature map of this module.
[0041] The deep feature extraction module in step 1 consists of five branches, each responsible for extracting features at different scales, thereby improving the network's ability to understand complex underwater scenes. The first branch consists of a regular convolution with a kernel of 1 and a SiLU activation function; the second branch consists of a regular convolution with a kernel of 1, a SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 3; the third branch consists of a pinwheel convolution with a kernel of 3, a SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 3; the fourth branch consists of a pinwheel convolution with a kernel of 5, a SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 5; the fifth branch is a short connection used to store the input features. The output features from the first to the fourth branch are fused through a concatenation operation. The fused feature map is then processed by a regular convolution and a SiLU activation function and added element-wise with the input features of this module to obtain the final output feature map.
[0042] The RFA module in step 1 consists of two branches. The first branch consists of average pooling, sigmoid function and CARAFFE module in sequence, and the output feature map of the first branch is E1. The second branch consists of windmill convolution with a kernel of 3, and its output feature map is E2. Feature map E1 is multiplied element by element, and the result is added element by element to feature map E1 to obtain feature map E3. Feature map E3 is adjusted by regular convolution to obtain the output feature map of RFA module.
[0043] In step 1, the four feature maps input into the improved feature fusion network are the output feature maps of the C1 module, C2 module, A2C2f1 module and RAF module, respectively. The A2C2f1 module is the A2C2f module in the original YOLOv12s.
[0044] In one embodiment of the present invention, the output feature map of module C3 in step 2 is subjected to a regular convolution with a kernel of 1 to obtain feature map D1; feature map D1 is split into feature map D2 and feature map D3 using a splitting operation, with a channel ratio of 1:1 between feature map D2 and feature map D3; feature map D2 is subjected to a sixth branch to obtain feature map D4, and feature map D2 is subjected to a seventh branch to obtain feature map D5; feature map D4, feature map D5, and feature map D6 are concatenated to obtain feature map D7; feature map D7 is subjected to a regular convolution to adjust the channel dimension, and the output feature map of module C3 is output; the sixth branch consists of two residual modules with the same structure, and the main module of each residual module is composed of a regular convolution with a kernel of 1, normalization, SiLU activation function, a windmill-shaped convolution with a kernel of 3, normalization, SiLU activation function, a regular convolution with a kernel of 1, normalization, and SiLU activation function in sequence. The deep feature map after RFA processing is upsampled by CARAFEE module 1 to output a feature map that retains deep semantics while recovering fine-grained localization information. This feature map is concatenated with the lateral output of A2C2f module 1 along the channel dimension. The output feature map has both deep and mid-level semantic features, which is beneficial for the detection and localization of both small and large targets. The concatenated feature map is then fed into A2C2f module 3. The output of A2C2f module 3 has a larger receptive field, unchanged resolution, enhanced local context, sufficient channel remixing, stable gradients, and is more sensitive to small / fine defects and better suppresses the background. The feature map output from A2C2f module 3 is upsampled by CARAFEE module 2 to obtain a feature map with higher resolution. This feature map is then concatenated with the lateral output of C2 module along the channel dimension to obtain a feature map that contains deep, mid, and shallow semantics. The feature map is input to A2C2f module 4, which outputs a feature map with higher resolution and deep, mid, and shallow semantics. The feature map output from A2C2f module 4 is upsampled by CARAFEE module 3 to obtain a higher resolution feature map. Simultaneously, it is concatenated with the lateral output of C1 module in the dimensional channel to obtain a feature map with deep, mid, and shallow semantics, and this feature map is more obvious in expressing the surface detail texture features of defects. The concatenated feature map is then passed to A2C2f module 5 to obtain a feature map with multiple scales and less noise. The lateral output of A2C2f module 5 is sent to detection head 4 to detect relatively small defects in communication optical cables. At the same time, the output of A2C2f module 5 is downsampled by convolutional layer 6 to obtain a smaller feature map. This feature map is concatenated with the lateral output of A2C2f module 4 in the channel dimension to obtain a multi-scale feature map with less noise. This feature map is then sent to A2C2f module 6 to obtain a feature map with a larger receptive field, preservation of edge details, and less background noise.Simultaneously, the lateral output of module A2C2f 6 is fed into detection head 3 to detect relatively small defects in the communication optical cable. The feature map output from module A2C2f 6 is downsampled by convolutional layer 7 to obtain a smaller feature map. This smaller feature map is then concatenated with the lateral output of module A2C2f 3 along the channel dimension. This concatenated feature map is then fed into module A2C2f 7 as input to obtain a feature map that is more responsive to medium-scale defects and has a better balance between detail and semantics. The lateral output feature map of module A2C2f 7 is then fed into detection head 2 to detect medium-sized defects in the communication optical cable. The output of module A2C2f 7 is downsampled by convolutional layer 8. The downsampled feature map is then concatenated with the lateral output of the RFA module along the channel dimension. This concatenated feature map is then fed into module C3. The feature map processed by module C3 is more sensitive to the overall shape and category of large / medium-scale defects, while suppressing large background textures and illumination interference, preserving geometric details, and enhancing semantics. Finally, the output feature map of module C3 is sent to detection head 1 to detect relatively large target defects in the communication optical cable.
[0045] In this embodiment, in order to verify the performance of the model disclosed in this invention, the YOLOv8s model, YOLOv10s model, YOLOv11s model, YOLOv12s model and the model of this patent were tested under the same conditions to perform defect detection performance tests on overhead communication optical cables. The results are shown in Table 1. The model disclosed in this patent is superior to other comparative models in terms of precision, recall and average accuracy.
[0046] Table 1
[0047]
[0048] Example 2
[0049] The present invention also provides a defect detection device for overhead power communication optical cables, comprising:
[0050] The first processing unit is used to input the inspection image of the overhead power communication optical cable into the improved backbone network and extract the features of the communication optical cable at different scales. The improved backbone network consists of convolutional layer 1, convolutional layer 2, C1 module, convolutional layer 3, C2 module, convolutional layer 4, A2C2f module 1, convolutional layer 5, A2C2f module 2, a depth feature extraction module, and an RFA module.
[0051] The second processing unit is used to output the feature information of four different scales from the first processing unit to the improved feature fusion network. The improved feature fusion network includes a first-layer fusion network and a second-layer fusion network. The first-layer fusion network consists of a CARAFEE module, a splicing operation, an A2C2f module 3, a CARAFEE module 2, a splicing operation, an A2C2f module 4, a CARAFEE module 3, and a splicing module in sequence. The second-layer fusion network consists of an A2C2f module 5, a convolutional layer 6, a convolutional layer 7, an A2C2f module 6, a convolutional layer 8, a splicing operation, an A2C2f module 7, a convolutional layer 9, a splicing operation, and a C3 module in sequence.
[0052] The third processing unit is used to input the four fused features output by the second processing unit into the detection network for detection, thereby realizing the detection of defects in power communication optical cables.
[0053] Example 3
[0054] The present invention also provides an overhead power communication optical cable defect detection system, comprising: a memory and a processor, wherein the memory stores a computer program executed by the processor, and the computer program executes an overhead power communication optical cable defect detection method when executed by the processor.
[0055] Example 4
[0056] The present invention also provides a storage medium storing a computer program, which executes a method for detecting defects in overhead power communication optical cables when running.
[0057] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for detecting defects in an aerial power communications optical cable, the method comprising: include: Step 1: Input the inspection images of overhead power communication optical cables into the improved backbone network and extract the features of communication optical cables at different scales; The improved backbone network consists of convolutional layer 1, convolutional layer 2, C1 module, convolutional layer 3, C2 module, convolutional layer 4, A2C2f module 1, convolutional layer 5, A2C2f module 2, deep feature extraction module, and RFA module in sequence. Step 2: Output the feature information of the four different scales from Step 1 to the improved feature fusion network; the improved feature fusion network includes a first-layer fusion network and a second-layer fusion network. The first-layer fusion network consists of a CARAFEE module, a stitching operation, an A2C2f module 3, a CARAFEE module 2, a stitching operation, an A2C2f module 4, a CARAFEE module 3, and a stitching module in sequence; the second-layer fusion network consists of an A2C2f module 5, a convolutional layer 6, a convolutional layer 7, an A2C2f module 6, a convolutional layer 8, a stitching operation, an A2C2f module 7, a convolutional layer 9, a stitching operation, and a C3 module in sequence. Step 3: Input the four fused features output from Step 2 into the detection network for detection, thereby realizing the detection of defects in power communication optical cables; The C1 and C2 modules in the improved backbone network have the same network structure. Their input feature maps are processed by a regular convolution with a kernel of 1 to adjust the number of channels of the input feature maps, resulting in feature map F1. After splitting feature map F1, feature maps F2 and F3 are obtained. Feature map F2 is then subjected to a regular convolution with kernel 1, normalization, SiLU activation, a depthwise separable convolution with kernel 3, normalization, SiLU activation, a regular convolution with kernel 1, normalization, and SiLU activation to obtain feature map F4. Feature map F4 is then subjected to a regular convolution with kernel 1 and SiLU activation to obtain the channel weights in feature map F4. These channel weights are multiplied by feature map F4 to obtain feature map F5. Feature map F5 is then subjected to a channel adjustment module consisting of a regular convolution with kernel 1, normalization, SiLU activation, and a regular convolution with kernel 1, and then element-wise added to feature map F4 to obtain feature map F6. Feature map F6 is concatenated with feature map F3 and feature map F1 to obtain feature map F7. Finally, feature map F7 is subjected to a regular convolution with kernel 1 to obtain C1. Output feature map of module or C2 module; The deep feature extraction module consists of five branches. The first branch consists of a regular convolution with a kernel of 1 and an SiLU activation function. The second branch consists of a regular convolution with a kernel of 1, an SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 3. The third branch consists of a pinwheel convolution with a kernel of 3, an SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 3. The fourth branch consists of a pinwheel convolution with a kernel of 5, an SiLU activation function, and a dilated convolution with a kernel of 3 and a dilation rate of 5. The fifth branch is a short connection used to store the input features. The output features from the first to the fourth branch are fused by a concatenation operation. The fused feature map is then subjected to regular convolution and SiLU activation function and added element-wise with the input features of the module to obtain the final output feature map. In step 2, the output feature map of module C3 is processed by a regular convolution with a kernel of 1 to obtain feature map D1. Feature map D1 is then split into feature map D2 and feature map D3 using a splitting operation, with a channel ratio of 1:1 between feature map D2 and feature map D3. Feature map D2 is processed by the sixth branch to obtain feature map D4, and by the seventh branch to obtain feature map D5. Feature maps D4, D5, and D6 are concatenated to obtain feature map D7. Feature map D7 is processed by a regular convolution to adjust the channel dimension, outputting the feature map of module C3. The sixth branch consists of two residual modules with identical structures. The main module of each residual module consists of, in sequence, a regular convolution with a kernel of 1, normalization, SiLU activation function, a windmill-shaped convolution with a kernel of 3, normalization, SiLU activation function, a regular convolution with a kernel of 1, normalization, and SiLU activation function.
2. The method for detecting defects in overhead power communication optical cables as described in claim 1, characterized in that, The RFA module consists of two branches. The first branch consists of average pooling, a sigmoid function, and a CARAFFE module, with the output feature map being E1. The second branch consists of a pinwheel convolution with a kernel of 3, with the output feature map being E2. Feature maps E1 and E2 are multiplied element-wise, and the result is added element-wise to E1 to obtain feature map E3. Feature map E3 is then adjusted using regular convolution to obtain the output feature map of the RFA module.
3. An overhead power communication optical cable defect detection device for implementing the defect detection method of claim 1, characterized in that, include: The first processing unit is used to input the inspection images of overhead power communication optical cables into the improved backbone network and extract the features of communication optical cables at different scales. The improved backbone network consists of convolutional layer 1, convolutional layer 2, C1 module, convolutional layer 3, C2 module, convolutional layer 4, A2C2f module 1, convolutional layer 5, A2C2f module 2, deep feature extraction module, and RFA module in sequence. The second processing unit is used to output the feature information of four different scales from the first processing unit to the improved feature fusion network. The improved feature fusion network includes a first-layer fusion network and a second-layer fusion network. The first-layer fusion network consists of a CARAFEE module, a concatenation operation, an A2C2f module 3, a CARAFEE module 2, a concatenation operation, an A2C2f module 4, a CARAFEE module 3, and a concatenation module in sequence. The second-layer fusion network consists of an A2C2f module 5, a convolutional layer 6, a convolutional layer 7, an A2C2f module 6, a convolutional layer 8, a concatenation operation, an A2C2f module 7, a convolutional layer 9, a concatenation operation, and a C3 module in sequence. The third processing unit is used to input the four fused features output by the second processing unit into the detection network for detection, thereby realizing the detection of defects in power communication optical cables.
4. A defect detection system for overhead power communication optical cables, characterized in that, include: A memory and a processor, wherein the memory stores a computer program executed by the processor, the computer program executing, when run by the processor, the method for detecting defects in overhead power communication optical cables as described in any one of claims 1-2.
5. A storage medium, characterized in that, The storage medium stores a computer program, which executes the overhead power communication optical cable defect detection method as described in any one of claims 1-2 when running.