Telegraph pole attached foreign matter detection method and device based on deep learning

Abstract

The invention discloses a telegraph pole attached foreign matter detection method and device based on deep learning, and is used for solving the technical problem that the final telegraph pole attached foreign matter detection effect is poor due to an existing telegraph pole attached foreign matter detection method. The method comprises the following steps: acquiring an image of the foreign matter attached to the telegraph pole, and inputting the image into a telegraph pole attached foreign matter detection network (including a multi-scale feature perception enhancement module, a separation weight cascade lightweight module, a double-roll multi-pooling small target detection module and a target detection head). A multi-scale feature perception enhancement module extracts features from the image to generate a multi-scale feature perception enhancement image; the image is input into a separation weight cascade lightweight module, and a separation weight cascade enhanced feature image is output; the double-roll multi-pooling small target detection module processes the foreign matter to generate a telegraph pole attached foreign matter detection feature map; and finally, the target detection head detects the feature map and outputs a detection result of the foreign matters attached to the telegraph pole.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method and device for detecting foreign objects attached to power poles based on deep learning. Background Technology

[0002] In the process of developing intelligent power systems, utility poles, as key infrastructure for power transmission, are prone to short circuits, tripping, and other faults caused by foreign objects such as plastic bags and kites attached to their surfaces, threatening the safe and stable operation of the power grid.

[0003] Traditional manual inspections are limited by efficiency and coverage, making them difficult to adapt to the real-time monitoring needs of large-scale power grids. Therefore, foreign object detection technology based on computer vision and deep learning has become a research focus. By acquiring data with image acquisition equipment and automatically identifying foreign objects through algorithms, inspection efficiency and fault early warning capabilities can be improved.

[0004] Most existing methods for detecting foreign objects attached to utility poles enhance features by adding attention modules or increasing the number of network layers. However, this significantly increases the number of network parameters and computational load, resulting in a significant decrease in inference speed. On the other hand, simplifying the network sacrifices its feature extraction capabilities, leading to a contradiction of "lightweight but inaccurate, accurate but inefficient," resulting in poor performance in detecting foreign objects attached to utility poles. Summary of the Invention

[0005] This invention provides a method and apparatus for detecting foreign objects attached to utility poles based on deep learning. It addresses the technical problem that existing methods for detecting foreign objects attached to utility poles suffer from a contradiction: strengthening features increases the number of parameters and computation, reduces inference speed, and simplifying the network results in a loss of feature extraction capabilities. This leads to a situation where "lightweight methods are inaccurate, while precise methods are inefficient," resulting in poor detection performance of foreign objects attached to utility poles.

[0006] The first aspect of this invention provides a method for detecting foreign objects attached to utility poles based on deep learning, comprising:

[0007] The image of foreign objects attached to a utility pole is acquired and input into a utility pole foreign object detection network. The utility pole foreign object detection network includes a multi-scale feature perception enhancement module, a separation weight cascade lightweight module, a dual-convolution multi-pooling small target detection module, and a target detection head.

[0008] The multi-scale feature perception enhancement module is used to extract features from the image of the foreign object attached to the utility pole to generate a multi-scale feature perception enhancement map.

[0009] The multi-scale feature perception enhancement map is used as the input of the separation weight cascade lightweight module, and the output is the separation weight cascade enhancement feature map;

[0010] The dual-convolution multi-pooling small target detection module performs dual-convolution multi-pooling small target detection on the separated weighted cascaded enhanced feature map to generate a feature map for detecting foreign objects attached to utility poles.

[0011] The target detection head is used to perform target detection on the feature map of foreign objects attached to the utility pole, and the detection result of foreign objects attached to the utility pole is output.

[0012] Optionally, the multi-scale feature perception enhancement module includes a 1×1 convolutional layer, a convolutional sub-module, a basic convolutional block, a dynamically deformable feature fusion sub-module, and a Sigmoid activation function layer; the step of using the multi-scale feature perception enhancement module to extract features from the image of the foreign object attached to the utility pole and generate a multi-scale feature perception enhancement map includes:

[0013] A 1×1 convolutional layer is used to perform a convolution operation on the image of the foreign object attached to the utility pole, outputting a first convolutional feature map of the utility pole. The first convolutional feature map of the utility pole is used as the input of the convolutional sub-module and the basic convolutional block, respectively, to output a second convolutional feature map of the utility pole and a third convolutional feature map of the utility pole.

[0014] The second utility pole convolutional feature map is dynamically deformable fused using the dynamic deformable feature fusion submodule to output the first utility pole dynamic deformable feature map.

[0015] The third utility pole convolutional feature map is dynamically deformable fused using the dynamic deformable feature fusion submodule to output the second utility pole dynamic deformable feature map.

[0016] The first dynamic deformable feature map of the utility pole is used as the input of the basic convolutional block to output the fourth convolutional feature map of the utility pole.

[0017] The second dynamic deformable feature map of the utility pole is used as the input of the convolutional submodule, and the fifth convolutional feature map of the utility pole is output.

[0018] The convolutional feature maps of the fifth and fourth utility poles are added element by element to output a convolutional spliced ​​feature map of the utility poles.

[0019] A 1×1 convolutional layer is used to perform a convolution operation on the convolutional splicing feature map of the utility pole, outputting a sixth utility pole convolutional feature map. The sixth utility pole convolutional feature map is then used as the input of a Sigmoid activation function layer to output a utility pole transformation feature map.

[0020] The transformed feature map of the utility pole and the convolutional feature map of the sixth utility pole are multiplied element-wise to output a multiplied feature map of the utility pole. The multiplied feature map of the utility pole is then convolved by a 1×1 convolutional layer to generate a multi-scale feature perception enhancement map.

[0021] Optionally, the dynamic deformable feature fusion submodule includes a 3×3 convolutional layer, a 7×7 convolutional layer, and a basic convolutional block; the step of performing dynamic deformable feature fusion on the second utility pole convolutional feature map through the dynamic deformable feature fusion submodule to output the first utility pole dynamic deformable feature map includes:

[0022] The second utility pole convolutional feature map is used as the input of the basic convolutional block to output the first utility pole basic convolutional feature map. The first utility pole basic convolutional feature map is then input into a 3×3 convolutional layer and a 7×7 convolutional layer for convolution operations to generate the seventh utility pole convolutional feature map and the eighth utility pole convolutional feature map.

[0023] Perform deformable convolution operation on the first utility pole base convolution feature map to output a utility pole deformable convolution feature map, and use the utility pole deformable convolution feature map as the input of the base convolution block to output a ninth utility pole convolution feature map.

[0024] The convolutional feature map of the seventh utility pole is downsampled to output a downsampled feature map of the utility pole, and the convolutional feature map of the eighth utility pole is upsampled to generate an upsampled feature map of the utility pole.

[0025] The downsampled feature map of the utility pole and the convolutional feature map of the ninth utility pole are concatenated to output the first concatenated feature map of the utility pole; and the upsampled feature map of the utility pole and the convolutional feature map of the ninth utility pole are concatenated to output the second concatenated feature map of the utility pole.

[0026] The first and second pole splicing feature maps are used as inputs to the basic convolutional block, and the first and second intermediate pole convolutional feature maps are output.

[0027] The first and second intermediate utility pole convolutional feature maps are added element by element to output the intermediate utility pole summed feature map. The intermediate utility pole summed feature map is then subjected to dynamic convolution operation to generate a utility pole dynamic convolutional feature map.

[0028] The dynamic convolutional feature map of the utility pole is used as the input to the basic convolutional block to output the first dynamic deformable feature map of the utility pole.

[0029] Optionally, the lightweight module with separated weights includes a 1×1 convolutional layer, a depthwise separable attention unit, a convolutional submodule, and a basic convolutional block; the step of using the multi-scale feature-aware enhancement map as input to the lightweight module with separated weights and outputting a separated weights-enhanced feature map includes:

[0030] The multi-scale feature perception enhancement map is convolved by a 1×1 convolutional layer to generate a first separation weight concatenated initial feature map, and the first separation weight concatenated initial feature map is enhanced and extracted by a depthwise separable attention unit to generate a second separation weight concatenated initial feature map.

[0031] The second separate weight concatenated initial feature map is used as the input of the convolutional submodule to output the third separate weight concatenated initial feature map. The third separate weight concatenated initial feature map is then enhanced and extracted using a depthwise separable attention unit to generate the fourth separate weight concatenated initial feature map.

[0032] The fourth separate weight concatenated initial feature map is used as the input of the basic convolutional block to output the fifth separate weight concatenated initial feature map. The fifth separate weight concatenated initial feature map is then enhanced and extracted through a depthwise separable attention unit to generate the sixth separate weight concatenated initial feature map.

[0033] The first separated weight concatenated initial feature map is used as the input of the convolutional submodule, and the seventh separated weight concatenated initial feature map is output.

[0034] The seventh separate weight concatenated initial feature map is used as the input to the basic convolutional block to output the eighth separate weight concatenated initial feature map;

[0035] The eighth and sixth initial feature maps of the separation weight concatenation are concatenated to output the ninth initial feature map of the separation weight concatenation. The ninth initial feature map of the separation weight concatenation is then convolved by a 1×1 convolutional layer to generate the enhanced feature map of the separation weight concatenation.

[0036] Optionally, the depthwise separable attention unit includes a basic convolutional block, a depthwise separable convolutional layer, and a spatial-channel attention mechanism layer; the step of performing feature enhancement extraction on the first separable weight concatenated initial feature map through the depthwise separable attention unit to generate a second separable weight concatenated initial feature map includes:

[0037] The first concatenated feature map of the separated weights is used as the input to the basic convolutional block, and the intermediate concatenated feature map of the separated weights is output.

[0038] A depthwise separable convolutional layer is used to perform depthwise convolution on the intermediate separation weight concatenated initial feature map to generate a first depthwise separable attention feature map. Max pooling and average pooling are then performed on the first depthwise separable attention feature map to output a second depthwise separable attention feature map and a third depthwise separable attention feature map.

[0039] Attention is calculated on the intermediate separation weight concatenated initial feature map through the spatial-channel attention mechanism layer, and a fourth depth separable attention feature map is output. The fourth depth separable attention feature map is then used as the input of the basic convolutional block to output a fifth depth separable attention feature map.

[0040] The second and third depthwise separable attention feature maps are used as inputs to the basic convolutional block, respectively, to output the sixth and seventh depthwise separable attention feature maps.

[0041] The sixth, seventh, and fifth depth-separable attention feature maps are concatenated to output the eighth depth-separable attention feature map.

[0042] The eighth deep separable attention feature map is used as the input of the basic convolutional block to generate the ninth deep separable attention feature map, and the spatial-channel attention mechanism layer is used to perform attention calculation on the ninth deep separable attention feature map to generate the tenth deep separable attention feature map.

[0043] The tenth depth separable attention feature map is used as the input to the basic convolutional block, and the second separable weight concatenated initial feature map is output.

[0044] Optionally, the dual-convolutional multi-pooling small target detection module includes a basic convolutional block, a convolutional sub-module, and a Sigmoid activation function layer; the step of performing dual-convolutional multi-pooling small target detection on the separated weighted cascaded enhanced feature map through the dual-convolutional multi-pooling small target detection module to generate a feature map for detecting foreign objects attached to utility poles includes:

[0045] The separated weighted concatenated enhanced feature map is used as the input of the basic convolutional block to output the first dual-convolution multi-pooling small target detection feature map, and the first dual-convolution multi-pooling small target detection feature map is subjected to dilated convolution to generate the second dual-convolution multi-pooling small target detection feature map.

[0046] Average pooling is performed on the second dual-convolution multi-pooling small target detection feature map to output the third dual-convolution multi-pooling small target detection feature map, and dynamic convolution is performed on the third dual-convolution multi-pooling small target detection feature map to generate the fourth dual-convolution multi-pooling small target detection feature map.

[0047] The fourth dual-convolution multi-pooling small target detection feature map is subjected to average pooling to output the fifth dual-convolution multi-pooling small target detection feature map;

[0048] Dynamic convolution is performed on the first dual-convolution multi-pooling small target detection feature map to generate a sixth dual-convolution multi-pooling small target detection feature map, and max pooling is performed on the sixth dual-convolution multi-pooling small target detection feature map to output a seventh dual-convolution multi-pooling small target detection feature map.

[0049] The seventh dual-convolution multi-pooling small target detection feature map is subjected to dilated convolution to generate the eighth dual-convolution multi-pooling small target detection feature map, and the eighth dual-convolution multi-pooling small target detection feature map is subjected to max pooling to output the ninth dual-convolution multi-pooling small target detection feature map.

[0050] The ninth dual-convolutional multi-pooling small target detection feature map and the fifth dual-convolutional multi-pooling small target detection feature map are added element-wise to generate the tenth dual-convolutional multi-pooling small target detection feature map. The tenth dual-convolutional multi-pooling small target detection feature map is used as the input of the basic convolutional block to output the eleventh dual-convolutional multi-pooling small target detection feature map.

[0051] The eleventh dual-volume multi-pooling small target detection feature map is used as the input to the Sigmoid activation function layer to output the intermediate dual-volume multi-pooling small target detection feature map. The intermediate dual-volume multi-pooling small target detection feature map and the tenth dual-volume multi-pooling small target detection feature map are multiplied element-wise to output the twelfth dual-volume multi-pooling small target detection feature map.

[0052] The twelfth dual-convolution multi-pooling small target detection feature map is used as the input of the convolution submodule to output the thirteenth dual-convolution multi-pooling small target detection feature map. The thirteenth dual-convolution multi-pooling small target detection feature map is then subjected to dilated convolution and dynamic convolution to generate the fourteenth dual-convolution multi-pooling small target detection feature map and the fifteenth dual-convolution multi-pooling small target detection feature map.

[0053] The fourteenth dual-volume multi-pooling small target detection feature map and the fifteenth dual-volume multi-pooling small target detection feature map are concatenated to output the sixteenth dual-volume multi-pooling small target detection feature map;

[0054] Using the sixteenth dual-convolution multi-pooling small target detection feature map as the input to the basic convolution block, the output is a feature map of foreign objects attached to utility poles.

[0055] A second aspect of the present invention provides a deep learning-based foreign object detection device for utility poles, comprising:

[0056] The acquisition module is used to acquire images of foreign objects attached to utility poles and input the images of foreign objects attached to utility poles into a foreign object detection network. The foreign object detection network includes a multi-scale feature perception enhancement module, a separation weight cascade lightweight module, a dual-convolution multi-pooling small target detection module, and a target detection head.

[0057] A module is used to extract features from the image of the foreign object attached to the utility pole using the multi-scale feature perception enhancement module, and generate a multi-scale feature perception enhancement map.

[0058] The input module is used to take the multi-scale feature perception enhancement map as the input of the separation weight cascade lightweight module and output the separation weight cascade enhancement feature map.

[0059] The first detection module is used to perform dual-convolution multi-pooling small target detection on the separated weighted cascaded enhanced feature map through the dual-convolution multi-pooling small target detection module to generate a feature map of foreign objects attached to utility poles.

[0060] The second detection module is used to perform target detection on the feature map of foreign objects attached to the utility pole using the target detection head, and output the detection result of foreign objects attached to the utility pole.

[0061] A computer device provided in a third aspect of the present invention includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the deep learning-based method for detecting foreign objects attached to utility poles as described in any of the preceding claims.

[0062] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the steps of the deep learning-based method for detecting foreign objects attached to utility poles as described in any of the preceding claims.

[0063] The fifth aspect of the present invention provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein, when the program instructions are executed by a computer, the computer performs the steps of the deep learning-based method for detecting foreign objects attached to utility poles as described in any of the preceding claims.

[0064] As can be seen from the above technical solutions, the present invention has the following advantages:

[0065] The present invention provides a deep learning-based method for detecting foreign objects attached to utility poles. The method acquires images of foreign objects attached to utility poles and inputs these images into a detection network. This network includes a multi-scale feature perception enhancement module, a lightweight separation weight cascade module, a dual-convolution multi-pooling small target detection module, and a target detection head. The multi-scale feature perception enhancement module extracts features from the foreign object images to generate a multi-scale feature perception enhancement map. This enhanced map is then used as input to the lightweight separation weight cascade module, which outputs a separate weight cascade enhancement feature map. Finally, the dual-convolution multi-pooling small target detection module further processes the enhanced feature map. A dual-convolution multi-pooling small target detection method is used to generate a feature map for detecting foreign objects attached to utility poles. A target detection head is then used to perform target detection on this feature map, outputting the detection results. Based on this approach, this invention specifically enhances feature extraction capabilities through a multi-scale feature perception enhancement module. Simultaneously, relying on a lightweight architecture design with a cascaded, weighted module, the number of network parameters and computational load are significantly reduced while ensuring feature enhancement effectiveness, effectively addressing the pain point of decreased inference speed when enhancing features in traditional methods. Furthermore, the dual-convolution multi-pooling small target detection module's accurate capture capability for small-sized foreign objects synergizes with the target detection head, thereby improving the detection effect of foreign objects attached to utility poles. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be 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.

[0067] Figure 1 The flowchart illustrates the steps of a method for detecting foreign objects attached to utility poles based on deep learning, as provided in Embodiment 1 of the present invention.

[0068] Figure 2 This is a structural block diagram of a foreign object detection network for utility poles provided in Embodiment 1 of the present invention;

[0069] Figure 3 This is a structural block diagram of the multi-scale feature perception enhancement module provided in Embodiment 1 of the present invention;

[0070] Figure 4 This is a structural block diagram of the dynamic deformable feature fusion submodule provided in Embodiment 1 of the present invention;

[0071] Figure 5 This is a structural block diagram of the cascaded lightweight module with separate weights provided in Embodiment 1 of the present invention;

[0072] Figure 6 This is a structural block diagram of the depth-separable attention unit provided in Embodiment 1 of the present invention;

[0073] Figure 7 This is a structural block diagram of the dual-convolution multi-pooling small target detection module provided in Embodiment 1 of the present invention;

[0074] Figure 8 This is a schematic diagram of an image of a foreign object attached to a utility pole, provided in Embodiment 1 of the present invention.

[0075] Figure 9 This is a schematic diagram of the detection results of foreign objects attached to a utility pole provided in Embodiment 1 of the present invention;

[0076] Figure 10 This is a flowchart illustrating a method for detecting foreign objects attached to utility poles based on deep learning, provided in Embodiment 1 of the present invention.

[0077] Figure 11 This is a structural block diagram of a foreign object detection device for utility poles based on deep learning, provided in Embodiment 2 of the present invention. Detailed Implementation

[0078] This invention provides a method and apparatus for detecting foreign objects attached to utility poles based on deep learning. It addresses the technical problem that existing methods for detecting foreign objects attached to utility poles suffer from a contradiction: strengthening features increases the number of parameters and computation, reduces inference speed, and simplifying the network results in a loss of feature extraction capabilities. This leads to a situation where "lightweight methods are inaccurate, while precise methods are inefficient," resulting in poor detection performance of foreign objects attached to utility poles.

[0079] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0080] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a method for detecting foreign objects attached to utility poles based on deep learning, as provided in Embodiment 1 of the present invention.

[0081] This invention provides a method for detecting foreign objects attached to utility poles based on deep learning, comprising:

[0082] Step 101: Obtain an image of foreign objects attached to a utility pole and input the image into the foreign object detection network for the utility pole. The foreign object detection network for the utility pole includes a multi-scale feature perception enhancement module, a separation weight cascade lightweight module, a dual-convolution multi-pooling small target detection module, and a target detection head.

[0083] It should be noted that the process involves using a photographic device to capture images of the utility pole with foreign objects attached to it. Specific photographic devices may include, but are not limited to, cameras, drones, and surveillance cameras.

[0084] Furthermore, this invention proposes the MSPFOD-Net (Multi-Module Synergy Pole Foreign Object Detection Network) for detecting foreign objects attached to power poles, overcoming existing technological bottlenecks. This network integrates three core modules: a multi-scale feature perception enhancement module dynamically fuses features from different levels and receptive fields to improve the recognition adaptability of large and small-scale foreign objects in complex scenes; a lightweight, weighted, cascaded module leverages depthwise separable convolution and attention mechanisms for synergistic optimization, reducing computational load while enhancing feature interaction through cascading, achieving a balance between lightweight deployment and high-precision detection; and a dual-convolution, multi-pooling small target detection module combines dilated convolution, dynamic convolution, and multi-pooling strategies to enhance the saliency of small target features and suppress background interference, solving the problems of missed and false detections of small foreign objects, and providing an efficient and accurate intelligent detection solution for power line inspection. The main structure of MSPFOD-Net is as follows: Figure 2 As shown,

[0085] exist Figure 2 In this context, Conv_BN_Mish (Convolution-Batch Normalization-Mish) represents a common neural network module, typically used in image processing, particularly in Convolutional Neural Networks (CNNs). It consists of three parts: a convolutional layer (Conv), a batch normalization layer (BN), and an activation function (Mish).

[0086] Concat is a feature fusion operation commonly used in deep learning models, widely applied due to its efficient structural design and low computational complexity. By concatenating multiple feature maps along the channel dimension, it integrates information from different layers or paths, enhancing the model's expressive power. Concat, which concatenates multiple convolutional feature maps along the channel dimension, is widely used in multi-path feature fusion and cross-layer feature connections, enhancing the model's expressive capabilities.

[0087] Dynamic Convolution (VCC) is a deep learning technique that breaks through the limitations of traditional static convolution. It departs from the fixed kernel parameters and dynamically generates kernel parameters based on the input data through attention mechanisms or weight fusion strategies. In actual operation, the model assigns different weights to each input sample and adaptively combines the basic convolution kernels to more accurately capture data features.

[0088] DilatedConv (dilated convolution) is a convolution operation that expands the receptive field by introducing "holes" (the spacing is controlled by the dilation rate) into the standard convolution kernel. It achieves an exponential increase in the coverage of the convolution kernel without increasing too many parameters and computational cost, which is significantly better than the linear growth of traditional convolution.

[0089] DWConv (Depth-wise Convolution) represents a depthwise separable convolution operation where the convolution kernel operates independently on each input channel, extracting spatial features from a single channel of the input feature map without involving cross-channel weight sharing. It preserves the unique information of each channel of the input feature map while significantly reducing computational cost and parameter count. It efficiently captures local patterns in data while reducing computational complexity.

[0090] DeformableConv (Deformable Convolution) is a convolution technique that adaptively adjusts the sampling position, breaking the limitation of fixed grid sampling in traditional convolution. By learning an additional set of offsets on the fixed sampling points of the standard convolution kernel, it enables the convolution kernel to dynamically adjust the sampling position according to the input features, thus more flexibly capturing irregularly shaped targets or complex textures.

[0091] MaxPool is a commonly used pooling method that reduces data dimensionality and extracts important features by selecting the maximum value within a local region.

[0092] AvgPool (Average Pooling) represents a pooling method used to reduce the spatial dimension of a feature map while preserving the average feature information of local regions.

[0093] CBS (Convolution-Batch Normalization-Swish) represents an efficient basic convolutional module that extracts local features through 3×3 convolutional kernels, stabilizes the distribution through Batch Normalization, and then introduces non-linearity to enhance expressive power through the Swish activation function. It is widely used in visual backbone networks, balancing lightweight design with feature enhancement.

[0094] CBAM (Convolutional Block Attention Module) represents a network that combines spatial and channel attention mechanisms to enhance the feature representation capabilities of convolutional neural networks (CNNs). The core idea of ​​CBAM is to allow the network to adaptively focus on more important features while suppressing irrelevant or redundant information.

[0095] The Head represents the object detection head, which is responsible for parsing feature maps and directly predicting the object's category, location, and confidence level. Its output is a three-dimensional tensor, where the prediction information for each spatial location includes the class probability of the predicted bounding box, the bounding box coordinates, and the object presence confidence level. These are key parameters for model localization and classification.

[0096] Step 102: Use a multi-scale feature perception enhancement module to extract features from the image of foreign objects attached to the utility pole and generate a multi-scale feature perception enhancement map.

[0097] The multi-scale feature perception enhancement module includes a 1×1 convolutional layer, a convolutional sub-module, a basic convolutional block, a dynamically deformable feature fusion sub-module, and a Sigmoid activation function layer.

[0098] It should be noted that the image of foreign objects attached to the utility pole is input into the multi-scale feature perception enhancement module to obtain a multi-scale feature perception enhancement map. In this module, the basic features of the image are first extracted, followed by dual-path differential processing. In the two different paths, different operation sequences are used to enhance and optimize the nonlinear expression of the features, leveraging the Dynamic-Deformable Feature Merge (DDFM) submodule to deeply mine complementary information. Subsequently, the dual-path outputs are fused, aggregating multi-dimensional cues to enhance the multi-scale features of the foreign objects attached to the utility pole. Then, sigmoid activation and element-wise multiplication are used to accurately filter out effective information highly correlated with the foreign objects. Finally, high-quality features adapted to downstream tasks are output, significantly improving the efficiency and accuracy of detecting foreign objects attached to utility poles. The structure of the multi-scale feature perception enhancement module is as follows: Figure 3 As shown, the multi-scale feature perception enhancement module contains two structurally identical dynamically deformable feature fusion sub-modules.

[0099] Furthermore, step 102 may include the following sub-steps:

[0100] S21. A 1×1 convolutional layer is used to perform a convolution operation on the image of the foreign object attached to the utility pole, outputting the first convolutional feature map of the utility pole. The first convolutional feature map of the utility pole is used as the input of the convolutional sub-module and the basic convolutional block, respectively, to output the second convolutional feature map of the utility pole and the third convolutional feature map of the utility pole.

[0101] It should be noted that the image of the foreign object attached to the utility pole is subjected to a Conv1x1 convolution operation to obtain a multi-scale feature perception enhancement map X1 (i.e., the first utility pole convolution feature map). X1 is then input into the Conv_BN_Mish submodule (convolution submodule) for operation, which enhances the features to give them a richer non-linear expression and captures the edge details of the foreign object attached to the utility pole, resulting in a multi-scale feature perception enhancement map X2 (i.e., the second utility pole convolution feature map). X1 is then input into the CBS submodule (basic convolution block) for operation, which optimizes the feature expression, resulting in a multi-scale feature perception enhancement map X5 (i.e., the third utility pole convolution feature map).

[0102] S22. The dynamic deformable feature fusion submodule is used to perform dynamic deformable feature fusion on the second utility pole convolutional feature map to output the first utility pole dynamic deformable feature map.

[0103] The dynamic deformable feature fusion submodule includes a 3×3 convolutional layer, a 7×7 convolutional layer, and a basic convolutional block.

[0104] It should be noted that inputting X2 into the DDFM submodule yields the multi-scale feature perception enhancement map X3 (i.e., the dynamic-deformable feature map of the first utility pole). The structure of the DDFM (Dynamic-Deformable Feature Merge) submodule within the multi-scale feature perception enhancement module is as follows: Figure 4 As shown.

[0105] Furthermore, step S22 may include the following sub-steps:

[0106] S221. The second utility pole convolution feature map is used as the input of the basic convolution block, the first utility pole basic convolution feature map is output, and the first utility pole basic convolution feature map is input into the 3×3 convolution layer and the 7×7 convolution layer respectively for convolution operation to generate the seventh utility pole convolution feature map and the eighth utility pole convolution feature map.

[0107] S222. Perform deformable convolution operation on the first utility pole base convolution feature map, output the utility pole deformable convolution feature map, and use the utility pole deformable convolution feature map as the input of the base convolution block to output the ninth utility pole convolution feature map.

[0108] S223. Downsample the convolutional feature map of the seventh utility pole to output the downsampled feature map of the utility pole, and upsample the convolutional feature map of the eighth utility pole to generate the upsampled feature map of the utility pole.

[0109] S224. Concatenate the downsampled feature map of the utility pole and the convolutional feature map of the ninth utility pole to output the first concatenated feature map of the utility pole, and concatenate the upsampled feature map of the utility pole and the convolutional feature map of the ninth utility pole to output the second concatenated feature map of the utility pole.

[0110] S225. Take the first and second pole splicing feature maps as inputs to the basic convolutional block, and output the first and second intermediate pole convolutional feature maps.

[0111] S226. Add the first and second intermediate utility pole convolution feature maps element by element to output the intermediate utility pole summed feature map, and perform dynamic convolution operation on the intermediate utility pole summed feature map to generate the utility pole dynamic convolution feature map.

[0112] S227. Use the dynamic convolutional feature map of the utility pole as the input of the basic convolutional block to output the first dynamic deformable feature map of the utility pole.

[0113] It should be noted that the multi-scale feature perception enhancement map X2 is input into the CBS submodule for processing to optimize the feature representation of foreign objects attached to the utility pole, resulting in the multi-scale feature perception enhancement map A1 (i.e., the first utility pole basic convolutional feature map). Subsequently, A1 is processed along three paths:

[0114] In path one, a Conv3×3 convolution operation is performed on A1. The 3×3 small convolution kernel is used to capture the local fine features of the foreign objects attached to the utility pole, enhance the expression of local details, and obtain the multi-scale feature perception enhancement map A2 (i.e., the convolution feature map of the seventh utility pole). Then, a Subsample operation is performed on A2 to reduce the resolution to highlight the overall features and compress the computation, and obtain the multi-scale feature perception enhancement map A3 (i.e., the downsampled feature map of the utility pole).

[0115] In path two, a Deformable Conv operation is performed on A1. By dynamically adjusting the sampling position of the convolution kernel, the deformation of the object in the feature map is adapted, and the non-rigid features of the foreign objects attached to the utility pole are accurately captured, resulting in a multi-scale feature perception enhancement map A4 (i.e., the deformable convolution feature map of the utility pole). A4 is then input into the CBS submodule for operation to optimize the feature expression and enhance the consistency and effectiveness of the features, resulting in a multi-scale feature perception enhancement map A5 (i.e., the convolution feature map of the ninth utility pole).

[0116] In path three, a Conv7×7 convolution operation is performed on A1. The 7×7 large convolution kernel covers a wider area and captures the global context features of the foreign objects attached to the utility pole, resulting in a multi-scale feature perception enhancement map A6 (i.e., the convolution feature map of the eighth utility pole). Then, an Upsample operation is performed on A6 to restore the spatial resolution of the feature map, resulting in a multi-scale feature perception enhancement map A7 (i.e., the upsampled feature map of the utility pole).

[0117] Then, A3 and A5 are concatenated using the Concat (channel splicing) operation to fuse local fine features and obtain a multi-scale feature perception enhancement map A8 (i.e., the first utility pole spliced ​​feature map). A8 is then input into the CBS submodule for operation to optimize feature representation and obtain a multi-scale feature perception enhancement map A9 (i.e., the first intermediate utility pole convolutional feature map).

[0118] Further, A7 and A5 are concatenated through the Concat (channel splicing) operation to fuse global context features and obtain a multi-scale feature perception enhancement map A10 (i.e., the second utility pole splicing feature map). Then, A10 is input into the CBS submodule for operation to optimize feature expression and obtain a multi-scale feature perception enhancement map A11 (i.e., the second middle utility pole convolution feature map).

[0119] Next, A9 and A11 are added element-wise to integrate feature information from different paths and fusion strategies, resulting in a multi-scale feature perception enhancement map A12 (i.e., the feature map of the middle utility pole). Then, a Dynamic Conv operation is performed on A12 to adaptively adjust the convolution kernel weights according to the feature content, accurately enhancing the key information of foreign objects attached to the utility pole and suppressing redundant background, resulting in a multi-scale feature perception enhancement map A13 (i.e., the dynamic convolution feature map of the utility pole).

[0120] Finally, A13 is input into the CBS submodule for operation to optimize feature representation, complete the final feature regularization, and obtain the multi-scale feature perception enhancement map X3 (i.e., the dynamic deformable feature map of the first utility pole).

[0121] The DDFM submodule utilizes three paths to comprehensively mine the multi-scale characteristics, deformation robustness, and contextual features of the input feature map. By processing different scales and feature forms through different paths, it effectively captures the characteristic performance of foreign objects attached to utility poles at different scales, enhancing adaptability to object deformation, and fully leveraging contextual features for accurate object identification. This processing significantly improves feature quality, making it better suited to the task requirements of complex scenarios such as detecting foreign objects attached to utility poles. It provides more comprehensive and accurate feature support for subsequent detection processes, achieving precise identification and localization of foreign objects attached to utility poles.

[0122] S23. The dynamic deformable feature fusion submodule is used to perform dynamic deformable feature fusion on the convolutional feature map of the third utility pole to output the dynamic deformable feature map of the second utility pole.

[0123] S24. Use the dynamic deformable feature map of the first utility pole as the input of the basic convolutional block, and output the convolutional feature map of the fourth utility pole.

[0124] S25. Use the dynamic deformable feature map of the second utility pole as the input of the convolutional submodule and output the convolutional feature map of the fifth utility pole.

[0125] S26. Add the convolutional feature maps of the fifth and fourth utility poles element by element to output the convolutional spliced ​​feature map of the utility poles.

[0126] S27. A 1×1 convolutional layer is used to perform a convolution operation on the convolutional splicing feature map of the utility pole, outputting the sixth utility pole convolutional feature map. The sixth utility pole convolutional feature map is then used as the input of the Sigmoid activation function layer to output the utility pole transformation feature map.

[0127] S28. Multiply the transformed feature map of the utility pole and the convolutional feature map of the sixth utility pole element by element to output the multiplied feature map of the utility pole. Then, perform a convolution operation on the multiplied feature map of the utility pole through a 1×1 convolutional layer to generate a multi-scale feature perception enhancement map.

[0128] It should be noted that X3 is input into the CBS submodule for operation to optimize the feature representation, resulting in the multi-scale feature perception enhancement map X4 (i.e., the fourth utility pole convolutional feature map). X5 is input into the DDFM submodule to obtain the multi-scale feature perception enhancement map X6 (i.e., the second utility pole dynamic deformable feature map). X6 is input into the Conv_BN_Mish submodule for operation to enhance the features, giving them a richer non-linear expression and capturing the edge details of foreign objects attached to the utility pole, resulting in the multi-scale feature perception enhancement map X7 (i.e., the fifth utility pole convolutional feature map).

[0129] Furthermore, X4 and X7 are added element-wise to fuse the dual-branch features. This process preserves the differentiated information of the two paths and allows the features to directly complement each other at the numerical level. This quickly aggregates multi-scale cues of foreign objects attached to utility poles, enhancing the integrity and discriminability of the features of foreign objects attached to utility poles, and yields the multi-scale feature perception enhancement map X8 (i.e., the convolutional splicing feature map of utility poles).

[0130] Furthermore, a Conv1x1 operation is performed on X8 to obtain a multi-scale feature perception enhancement map X9 (i.e., the sixth utility pole convolution feature map). Subsequently, X9 is processed using the Sigmoid activation function, and then its output is multiplied with X9 element-wise to automatically suppress background areas (such as sky and vegetation) and accurately locate the position of foreign objects attached to the utility pole, resulting in a multi-scale feature perception enhancement map X10 (i.e., the utility pole multiplication feature map).

[0131] Finally, Conv1x1 operation is performed on X10 to obtain the multi-scale feature perception enhancement map X11 (i.e., the multi-scale feature perception enhancement map).

[0132] It is worth mentioning that the processing process of the dynamic deformable feature map of the second utility pole is the same as that of the dynamic deformable feature map of the first utility pole, and will not be described in detail here.

[0133] Step 103: Use the multi-scale feature perception enhancement map as input to the separation weight cascade lightweight module, and output the separation weight cascade enhancement feature map.

[0134] The separated weighted cascaded lightweight module includes a 1×1 convolutional layer, a depthwise separable attention unit, a convolutional submodule, and a basic convolutional block.

[0135] It should be noted that the multi-scale feature perception enhancement map is input into the lightweight module of the separation weight cascade to obtain the separation weight cascade enhanced feature map. The lightweight module of the separation weight cascade employs a dual-path design: in path one, multi-level DSAU is used to progressively and deeply mine feature associations, accurately extracting key features of foreign objects attached to utility poles, resulting in richer feature semantics and facilitating accurate identification of foreign object features; in path two, lightweight operations are used to quickly supplement basic features, constructing complementary information, enriching feature dimensions while ensuring efficiency. After the dual-path features are spliced, fused, and dimensionally adjusted, the output features retain rich semantics while possessing computational efficiency. This ensures feature quality while precisely meeting the computational cost control requirements of lightweight networks, providing strong support for detecting foreign objects attached to utility poles, enabling more efficient and accurate detection of such objects. The structure of the lightweight module of the separation weight cascade is as follows: Figure 5 As shown, the separable weight cascaded lightweight module contains three structurally identical depth-separable attention units (DSAUs).

[0136] Furthermore, step 103 may include the following sub-steps:

[0137] S31. A 1×1 convolutional layer is used to perform a convolution operation on the multi-scale feature perception enhancement map to generate the first separation weight concatenation initial feature map. The first separation weight concatenation initial feature map is then enhanced and extracted using a depthwise separable attention unit to generate the second separation weight concatenation initial feature map.

[0138] The depthwise separable attention unit includes a basic convolutional block, a depthwise separable convolutional layer, and a spatial-channel attention mechanism layer.

[0139] It should be noted that the multi-scale feature perception enhancement map X11 is subjected to a Conv1×1 convolution operation to obtain the separation weight concatenated enhancement feature map Y1 (i.e., the first separation weight concatenated initial feature map). Y1 is input into the DSAU submodule to adaptively focus on the key region of the foreign object attached to the utility pole, suppress background noise, and improve the feature response, resulting in the separation weight concatenated enhancement feature map Y2 (i.e., the second separation weight concatenated initial feature map). The structure of the DSAU (Depth-Separable Attention Unit) submodule is as follows: Figure 6 As shown.

[0140] Further, step S31 may include the following sub-steps:

[0141] S311. Use the first concatenated feature map of the separation weights as the input of the basic convolutional block, and output the intermediate concatenated feature map of the separation weights.

[0142] S312. A depthwise separable convolutional layer is used to perform depthwise convolution on the initial feature map of the intermediate separation weight concatenation to generate a first depthwise separable attention feature map. Max pooling and average pooling are then performed on the first depthwise separable attention feature map to output a second depthwise separable attention feature map and a third depthwise separable attention feature map.

[0143] S313. The spatial-channel attention mechanism layer performs attention calculation on the initial feature map of the intermediate separation weight concatenation, outputting the fourth depth separable attention feature map, and uses the fourth depth separable attention feature map as the input of the basic convolution block to output the fifth depth separable attention feature map.

[0144] S314. Take the second and third depthwise separable attention feature maps as inputs to the basic convolutional block, and output the sixth and seventh depthwise separable attention feature maps respectively.

[0145] S315. Concatenate the sixth, seventh, and fifth depth separable attention feature maps to output the eighth depth separable attention feature map.

[0146] S316. The eighth depth separable attention feature map is used as the input of the basic convolutional block to generate the ninth depth separable attention feature map. The spatial-channel attention mechanism layer is used to perform attention calculation on the ninth depth separable attention feature map to generate the tenth depth separable attention feature map.

[0147] S317. Take the tenth depth separable attention feature map as the input of the basic convolutional block and output the second separable weight concatenation initial feature map.

[0148] It should be noted that Y1 is input into the CBS submodule for processing to optimize the feature representation, resulting in the separated weight concatenated enhanced feature map B1 (i.e., the initial feature map of the intermediate separated weight concatenation). Subsequently, B1 is processed through three paths:

[0149] In Path 1, DWConv (Depthly Separable Convolution) is performed on B1 to capture local spatial features. Under the premise of lightweighting, fine details of foreign objects attached to the utility pole are extracted, resulting in a separated weight concatenated enhanced feature map B2 (i.e., the first depthly separable attention feature map). Then, MaxPool is performed on B2 to highlight the most significant local information in the feature map, strengthen the expression of key details, and reduce the spatial dimension to quickly locate the key position of the foreign objects attached to the utility pole, resulting in a separated weight concatenated enhanced feature map B3 (i.e., the second depthly separable attention feature map). Next, B3 is input into the CBS submodule for operation to optimize the feature expression, making the local detail features more suitable for subsequent fusion and improving the utilization efficiency of the detail features of the foreign objects attached to the utility pole, resulting in a separated weight concatenated enhanced feature map B4 (i.e., the sixth depthly separable attention feature map).

[0150] In path two, CBAM (spatial-channel attention mechanism layer) is applied to B1. Through the cascading of channel attention and spatial attention, key channel information is first focused, and then important spatial regions are strengthened to accurately mine the implicit association patterns in the feature map, so as to better identify the spatial relationship between the background and the foreign objects attached to the utility pole, resulting in the separation weight cascaded enhanced feature map B5 (i.e., the fourth depth separable attention feature map). Then, B5 is input into the CBS submodule for operation to optimize the feature expression, so that the association information between the channel and spatial dimensions is more stable and clearer, and the overall feature grasp of the foreign objects attached to the utility pole is improved, resulting in the separation weight cascaded enhanced feature map B6 (i.e., the fifth depth separable attention feature map).

[0151] In path three: DWConv (depth-separable convolution) is performed on B1 to extract spatial features in a lightweight manner, resulting in the separable weight concatenated enhanced feature map B7; then AvgPool (average pooling) is performed on B7 to smooth the feature map information, highlight global statistical regularities, and supplement the large-scale contextual features of foreign objects attached to utility poles, resulting in the separable weight concatenated enhanced feature map B8 (i.e., the third depth-separable attention feature map); then B8 is input into the CBS submodule for operation to optimize feature expression, regularize the global features after average pooling, eliminate pooling noise, make the global contextual features more adaptable and integrated, and improve the overall detection capability of foreign objects attached to utility poles, resulting in the separable weight concatenated enhanced feature map B9 (i.e., the seventh depth-separable attention feature map).

[0152] Then, B4, B6, and B9 are concat (channel splicing) to fuse multi-dimensional features such as local details, feature associations, and global context, comprehensively integrating the feature information of foreign objects attached to the utility pole, resulting in the separation weight concatenated enhanced feature map B10 (i.e., the eighth depth separable attention feature map). B10 is input into the CBS submodule for operation to optimize feature representation, resulting in the separation weight concatenated enhanced feature map B11 (i.e., the ninth depth separable attention feature map). Next, CBAM is applied to B11 to further enhance feature associations, accurately focus on key information, and suppress redundant content, resulting in the separation weight concatenated enhanced feature map B12 (i.e., the tenth depth separable attention feature map).

[0153] Finally, B12 is input into the CBS submodule for operation to optimize the feature representation and obtain the separated weight concatenated enhanced feature map Y2 (i.e. the second separated weight concatenated initial feature map).

[0154] The DSAU submodule employs a three-path differentiated extraction strategy in the scenario of detecting foreign objects attached to utility poles. It focuses on local details, feature associations, and global context, respectively, to comprehensively capture the feature information of the utility pole and the attached foreign objects. Simultaneously, through multiple rounds of CBS regularization operations, the feature structure is continuously optimized. Combined with the CBAM attention enhancement mechanism, key features are accurately highlighted, enabling the network to fully exploit the diverse information in the input feature map. This significantly improves feature quality and provides more comprehensive and accurate feature support for subsequent detection processes. It effectively enhances the network's ability to express and utilize diverse features related to foreign objects on utility poles, contributing to more accurate detection.

[0155] S32. The second separate weight concatenated initial feature map is used as the input of the convolutional sub-module to output the third separate weight concatenated initial feature map. The depthwise separable attention unit is used to perform feature enhancement extraction on the third separate weight concatenated initial feature map to generate the fourth separate weight concatenated initial feature map.

[0156] S33. The fourth separation weight concatenated initial feature map is used as the input of the basic convolutional block to output the fifth separation weight concatenated initial feature map. The fifth separation weight concatenated initial feature map is then enhanced and extracted through a depthwise separable attention unit to generate the sixth separation weight concatenated initial feature map.

[0157] S34. Use the first separated weight concatenated initial feature map as the input of the convolutional submodule and output the seventh separated weight concatenated initial feature map.

[0158] S35. Use the seventh separate weight concatenated initial feature map as the input to the basic convolutional block, and output the eighth separate weight concatenated initial feature map;

[0159] S36. Concatenate the eighth and sixth initial feature maps of the separation weight concatenation to output the ninth initial feature map of the separation weight concatenation. Then, perform a convolution operation on the ninth initial feature map of the separation weight concatenation through a 1×1 convolutional layer to generate the separation weight concatenation enhanced feature map.

[0160] It should be noted that Y2 is input into the Conv_BN_Mish submodule (i.e., the convolution submodule) for operation, which enhances the features to give them a richer nonlinear expression and highlights the edge details of the foreign objects attached to the utility pole, resulting in the separation weight concatenated enhanced feature map Y3 (i.e., the third separation weight concatenated initial feature map); then Y3 is input into the DSAU submodule to obtain the separation weight concatenated enhanced feature map Y4 (i.e., the fourth separation weight concatenated initial feature map); Y4 is input into the CBS submodule for operation and optimization of the feature expression, resulting in the separation weight concatenated enhanced feature map Y5 (i.e., the fifth separation weight concatenated initial feature map); finally, Y5 is input into the DSAU submodule to obtain the separation weight concatenated enhanced feature map Y6 (i.e., the sixth separation weight concatenated initial feature map).

[0161] Furthermore, Y1 is input into the Conv_BN_Mish submodule for processing, enhancing the features to give them a richer nonlinear expression and highlighting the edge details of foreign objects attached to the utility pole, resulting in the separated weight cascaded enhanced feature map Y7 (i.e., the seventh separated weight cascaded initial feature map). Y7 is then input into the CBS submodule for processing to optimize the feature expression, resulting in the separated weight cascaded enhanced feature map Y8 (i.e., the eighth separated weight cascaded initial feature map). Through lightweight operations, key information is supplemented with less computation while retaining basic features.

[0162] Furthermore, Y6 and Y8 are concat (channel splicing) to merge the differential features of the two paths, combine the advantages of deep enhancement and concise transmission, improve the recall rate of foreign objects attached to utility poles, and obtain the separated weight cascaded enhanced feature map Y9 (i.e. the ninth separated weight cascaded initial feature map).

[0163] Furthermore, a Conv1×1 convolution operation is performed on Y9, and the channel dimensions are adjusted again to obtain the separated weight concatenated enhanced feature map Y10 (i.e., the separated weight concatenated enhanced feature map).

[0164] It is worth mentioning that the processing procedures for the fourth and sixth separation weight concatenation initial feature maps are the same as those for the second separation weight concatenation initial feature map, and will not be elaborated further in this invention.

[0165] Step 104: Perform dual-convolution multi-pooling small target detection on the separated weighted concatenated enhanced feature map using the dual-convolution multi-pooling small target detection module to generate a feature map for detecting foreign objects attached to utility poles.

[0166] The dual-convolutional multi-pooling small target detection module includes a basic convolutional block, a convolutional sub-module, and a Sigmoid activation function layer.

[0167] It should be noted that the separated weighted feature map is input into the dual-convolution multi-pooling small target detection module to obtain the feature map for detecting foreign objects attached to utility poles. In this module, a dual-path differential extraction strategy is employed. One path deeply mines global features to comprehensively grasp the overall situation of the utility pole and its surrounding environment; the other path focuses on significant local features to accurately capture the details of the foreign objects attached to the pole. Subsequently, through cross-path fusion and attention mechanisms, key features related to the foreign objects attached to the pole are precisely enhanced, effectively filtering out irrelevant interference. Further, through secondary branch convolution and fusion, long-distance feature associations are deepened, allowing feature information from different locations to fully interact. Throughout the processing, unique patterns are independently enhanced, highlighting the individual features of the utility pole and the attached foreign objects, while simultaneously improving overall expressiveness, making the features more complete and accurate. The output features can accurately empower subsequent tasks related to detecting foreign objects attached to utility poles, providing richer and more discriminative support, adapting to feature enhancement needs in various scenarios, and achieving efficient and accurate detection of foreign objects attached to utility poles. The structure of the dual-convolution multi-pooling small target detection module is as follows: Figure 7 As shown.

[0168] Furthermore, step 104 may include the following sub-steps:

[0169] S41. Take the concatenated weighted enhanced feature map as the input of the basic convolutional block, output the first double-convolution multi-pooling small target detection feature map, and perform dilation convolution on the first double-convolution multi-pooling small target detection feature map to generate the second double-convolution multi-pooling small target detection feature map.

[0170] S42. Perform average pooling on the second dual-convolution multi-pooling small target detection feature map to output the third dual-convolution multi-pooling small target detection feature map, and perform dynamic convolution on the third dual-convolution multi-pooling small target detection feature map to generate the fourth dual-convolution multi-pooling small target detection feature map.

[0171] S43. Perform average pooling on the fourth dual-convolution multi-pooling small target detection feature map to output the fifth dual-convolution multi-pooling small target detection feature map;

[0172] S44. Perform dynamic convolution on the first dual-convolution multi-pooling small target detection feature map to generate the sixth dual-convolution multi-pooling small target detection feature map, and perform max pooling on the sixth dual-convolution multi-pooling small target detection feature map to output the seventh dual-convolution multi-pooling small target detection feature map.

[0173] S45. Perform dilated convolution on the seventh dual-convolution multi-pooling small target detection feature map to generate the eighth dual-convolution multi-pooling small target detection feature map, and perform max pooling on the eighth dual-convolution multi-pooling small target detection feature map to output the ninth dual-convolution multi-pooling small target detection feature map.

[0174] S46. Add the ninth dual-convolutional multi-pooling small target detection feature map and the fifth dual-convolutional multi-pooling small target detection feature map element by element to generate the tenth dual-convolutional multi-pooling small target detection feature map. Use the tenth dual-convolutional multi-pooling small target detection feature map as the input of the basic convolutional block and output the eleventh dual-convolutional multi-pooling small target detection feature map.

[0175] S47. The eleventh double-volume multi-pooling small target detection feature map is used as the input of the Sigmoid activation function layer, and the middle double-volume multi-pooling small target detection feature map is output. The middle double-volume multi-pooling small target detection feature map and the tenth double-volume multi-pooling small target detection feature map are multiplied element-wise to output the twelfth double-volume multi-pooling small target detection feature map.

[0176] S48. Take the twelfth double-pooled multi-pooled small target detection feature map as the input of the convolutional sub-module, output the thirteenth double-pooled multi-pooled small target detection feature map, and perform dilated convolution and dynamic convolution on the thirteenth double-pooled multi-pooled small target detection feature map respectively to generate the fourteenth double-pooled multi-pooled small target detection feature map and the fifteenth double-pooled multi-pooled small target detection feature map.

[0177] S49. The fourteenth dual-volume multi-pooling small target detection feature map and the fifteenth dual-volume multi-pooling small target detection feature map are concatenated to output the sixteenth dual-volume multi-pooling small target detection feature map;

[0178] S410. Use the sixteenth dual-convolution multi-pooling small target detection feature map as the input of the basic convolution block, and output the feature map of foreign objects attached to the utility pole.

[0179] It should be noted that Y10 is input into the CBS submodule for operation to optimize the feature representation, resulting in the feature map Z1 for detecting foreign objects attached to utility poles (i.e., the first dual-convolution multi-pooling small target detection feature map). Subsequently, Z1 is processed in two paths:

[0180] In path one, a Dilated Conv operation is performed on Z1 to increase the receptive field and capture long-range global features of foreign objects attached to utility poles without increasing the number of parameters, resulting in the utility pole attachment detection feature map Z2 (i.e., the second dual-convolution multi-pooling small target detection feature map). An AvgPool operation is then performed on Z2 to smooth the feature map information, highlight global statistical regularities, and supplement the large-scale contextual features of the foreign objects attached to the utility poles, resulting in the utility pole attachment detection feature map Z3 (i.e., the third dual-convolution multi-pooling small target detection feature map). Dynamic convolution is then performed on Z3. The Conv (Dynamic Convolution) operation adaptively adjusts the weights of the convolution kernel based on the feature content, accurately enhancing key information from a global perspective, suppressing redundant background, and accurately identifying the features of foreign objects attached to utility poles, resulting in the utility pole attachment foreign object detection feature map Z4 (i.e., the fourth dual-convolution multi-pooling small target detection feature map). The AvgPool operation is then performed on Z4 to further aggregate global features, compress spatial dimensions, and make the features more focused on core patterns, improving the overall detection effect of foreign objects attached to utility poles, resulting in the utility pole attachment foreign object detection feature map Z5 (i.e., the fifth dual-convolution multi-pooling small target detection feature map).

[0181] In path two, a Dynamic Conv operation is performed on Z1 to prioritize capturing local salient features. Key details are enhanced through adaptive weighting, accurately identifying the local features of foreign objects attached to the utility pole, resulting in the utility pole attachment detection feature map Z6 (i.e., the sixth dual-convolution multi-pooling small target detection feature map). A MaxPool operation is then performed on Z6 to highlight the most salient local features in the feature map, enhancing the expression of key details while reducing spatial dimensions, quickly locating the key positions of foreign objects attached to the utility pole, resulting in the utility pole attachment detection feature map Z7 (i.e., the seventh dual-convolution multi-pooling small target detection feature map). Finally, Z7 is dilated... The Conv (dilated convolution) operation expands the receptive field and associates long-range information on the basis of strengthening local features, allowing local details to be integrated with the global scene, resulting in a more comprehensive and accurate detection of foreign objects attached to utility poles. This yields the utility pole attachment detection feature map Z8 (i.e., the eighth dual-convolution multi-pooling small target detection feature map). The MaxPool operation is then performed on Z8 to further aggregate significant local features and compress the spatial dimension, resulting in the utility pole attachment detection feature map Z9 (i.e., the ninth dual-convolution multi-pooling small target detection feature map).

[0182] Furthermore, Z5 and Z9 are added element-wise to integrate global aggregation features and significant local aggregation features, allowing the advantages of different paths to complement each other at the numerical level. This fully integrates the features of foreign objects attached to utility poles, improving the accuracy and comprehensiveness of detection, and yields the feature map Z10 for foreign objects attached to utility poles (i.e., the tenth dual-volume multi-pooling small target detection feature map).

[0183] Furthermore, Z10 is input into the CBS submodule for operation to optimize the feature representation, resulting in the feature map Z11 for detecting foreign objects attached to utility poles (i.e., the eleventh dual-volume multi-pooling small target detection feature map). Subsequently, Z11 is processed using the Sigmoid activation function, and then its output is multiplied element-wise with Z11 to enhance the key features of foreign objects attached to utility poles and suppress irrelevant information, resulting in the feature map Z12 for detecting foreign objects attached to utility poles (i.e., the twelfth dual-volume multi-pooling small target detection feature map).

[0184] Furthermore, Z12 is input into the Conv_BN_Mish submodule for operation to obtain the feature map Z13 for detecting foreign objects attached to utility poles (i.e., the thirteenth dual-convolution multi-pooling small target detection feature map). Then, Z13 is subjected to Dilated Conv and Dynamic Conv operations to obtain the feature maps Z14 (i.e., the fourteenth dual-convolution multi-pooling small target detection feature map) and Z15 (i.e., the fifteenth dual-convolution multi-pooling small target detection feature map) for detecting foreign objects attached to utility poles. By using dual-branch dilated convolution to further mine long-distance feature associations, the long-distance relationship between foreign objects attached to utility poles and the surrounding environment can be analyzed more deeply, thus improving the detection effect.

[0185] Furthermore, the Concat (channel splicing) operation is performed on Z14 and Z15 to fuse the features after the dual-branch dilated convolution, resulting in the feature map Z16 for detecting foreign objects attached to utility poles (i.e., the sixteenth dual-convolution multi-pooling small target detection feature map). Z16 is then input into the CBS submodule for operation to optimize the feature representation, resulting in the feature map Z17 for detecting foreign objects attached to utility poles (i.e., the feature map for detecting foreign objects attached to utility poles).

[0186] Step 105: Use the target detection head to perform target detection on the feature map of foreign objects attached to the utility pole, and output the detection results of foreign objects attached to the utility pole.

[0187] It should be noted that by inputting the feature map of foreign objects attached to the utility pole into the Head, the detection results of foreign objects attached to the utility pole are obtained. The detection results include the bounding box coordinates and the category of foreign objects attached to the utility pole.

[0188] For example, a camera is used to photograph a utility pole to obtain an image of a foreign object attached to the pole, such as... Figure 8 As shown, the image of a utility pole with attached foreign objects is input into the multi-scale feature perception enhancement module to obtain a multi-scale feature perception enhancement map. An input image of a utility pole with attached foreign objects of size 1088×800×3 is subjected to a Conv1x1 convolution operation to obtain a multi-scale feature perception enhancement map X1 of size 1088×800×16. Then, feature processing is performed on X1 through two paths:

[0189] (1) In path one, X1 is input into the Conv_BN_Mish submodule for operation to obtain a multi-scale feature perception enhancement map X2 with a size of 1088×800×16; X2 is input into the DDFM submodule to obtain a multi-scale feature perception enhancement map X3 with a size of 544×400×32; X3 is then input into the CBS submodule for operation to obtain a multi-scale feature perception enhancement map X4 with a size of 544×400×32.

[0190] (2) In path two, X1 is input into the CBS submodule for operation to obtain a multi-scale feature perception enhancement map X5 with a size of 1088×800×16; X5 is input into the DDFM submodule to obtain a multi-scale feature perception enhancement map X6 with a size of 544×400×32; X6 is input into the Conv_BN_Mish submodule for operation to obtain a multi-scale feature perception enhancement map X7 with a size of 544×400×32.

[0191] Further, X4 and X7 are added element-wise to obtain a multi-scale feature perception enhancement map X8 with a size of 544×400×32. A Conv1x1 operation is performed on X8 to obtain a multi-scale feature perception enhancement map X9 with a size of 272×200×64. X9 is processed using the Sigmoid activation function, and its output is multiplied element-wise with X9 to obtain a multi-scale feature perception enhancement map X10 with a size of 272×200×64. A Conv1x1 operation is performed on X10 to obtain a multi-scale feature perception enhancement map X11 with a size of 272×200×64. The construction and operation process of the DDFM submodule is as follows: First, the multi-scale feature perception enhancement map X2 is input into the CBS submodule for operation to obtain a multi-scale feature perception enhancement map A1 with a size of 1088×800×16; then A1 is processed through three paths:

[0192] (1) In path one, perform Conv3×3 convolution operation on A1 to obtain a multi-scale feature perception enhancement map A2 with a size of 1088×800×16; perform Subsample operation on A2 to obtain a multi-scale feature perception enhancement map A3 with a size of 544×400×16.

[0193] (2) In path two, perform Deformable Conv operation on A1 to obtain a multi-scale feature perception enhancement map A4 with a size of 1088×800×16; input A4 into the CBS submodule for operation to obtain a multi-scale feature perception enhancement map A5 with a size of 1088×800×16.

[0194] (3) In path three, perform Conv7×7 convolution operation on A1 to obtain a multi-scale feature perception enhancement map A6 with a size of 1088×800×16; perform Upsample operation on A6 to obtain a multi-scale feature perception enhancement map A7 with a size of 2176×1600×16.

[0195] Then, A3 and A5 are concatenated using the Concat operation to obtain a multi-scale feature perception enhancement image A8 with a size of 544×400×32. A8 is then input into the CBS submodule for processing to obtain a multi-scale feature perception enhancement image A9 with a size of 544×400×32. A7 and A5 are then concatenated using the Concat operation to obtain a multi-scale feature perception enhancement image A10 with a size of 1088×800×32. A10 is then input into the CBS submodule for processing to obtain a multi-scale feature perception enhancement image A11 with a size of 1088×800×32. A9 and A11 are then element-wise added to obtain a multi-scale feature perception enhancement image A12 with a size of 544×400×32. A12 is then subjected to Dynamic... The Conv operation yields a multi-scale feature perception enhancement map A13 with a size of 544×400×32; A13 is then input into the CBS submodule for further processing to obtain a multi-scale feature perception enhancement map X3 with a size of 544×400×32.

[0196] Furthermore, the multi-scale feature-aware enhancement map is input into the lightweight module of the separation weight concatenation, resulting in a separation weight concatenation enhancement feature map. The multi-scale feature-aware enhancement map X11 is then subjected to a Conv1×1 convolution operation to obtain a separation weight concatenation enhancement feature map Y1 of size 272×200×64. Subsequently, Y1 is processed along two paths:

[0197] (1) In path one, Y1 is input into the DSAU submodule to obtain a cascaded enhancement feature map Y2 with a size of 272×200×64; Y2 is input into the Conv_BN_Mish submodule to obtain a cascaded enhancement feature map Y3 with a size of 272×200×64; Y3 is input into the DSAU submodule to obtain a cascaded enhancement feature map Y4 with a size of 272×200×64; Y4 is input into the CBS submodule to obtain a cascaded enhancement feature map Y5 with a size of 272×200×64; Y5 is input into the DSAU submodule to obtain a cascaded enhancement feature map Y6 with a size of 272×200×64.

[0198] (2) In path two, Y1 is input into the Conv_BN_Mish submodule for operation to obtain a separated weight concatenated enhancement feature map Y7 with a size of 272×200×64; Y7 is input into the CBS submodule for operation to obtain a separated weight concatenated enhancement feature map Y8 with a size of 272×200×64.

[0199] Further, Y6 and Y8 are concatenated to obtain a 272×200×128 separated weight concatenated enhanced feature map Y9; Y9 is then convolved with a Conv1×1 operation to obtain a 136×100×128 separated weight concatenated enhanced feature map Y10. The construction and execution process of the DSAU submodule is as follows: First, Y1 is input into the CBS submodule for processing, resulting in a 272×200×64 separated weight concatenated enhanced feature map B1; then, B1 is processed along three paths:

[0200] (1) In path one, perform DWConv operation on B1 to obtain a separated weight concatenated enhancement feature map B2 with a size of 272×200×64; perform MaxPool operation on B2 to obtain a separated weight concatenated enhancement feature map B3 with a size of 136×100×64; input B3 into the CBS submodule for operation to obtain a separated weight concatenated enhancement feature map B4 with a size of 136×100×64.

[0201] (2) In path two, apply CBAM to B1 to obtain a separated weight cascaded enhancement feature map B5 with a size of 272×200×64; input B5 into the CBS submodule for operation to obtain a separated weight cascaded enhancement feature map B6 with a size of 272×200×64.

[0202] (3) In path three, perform DWConv operation on B1 to obtain a separated weight concatenated enhancement feature map B7 with a size of 272×200×64; perform AvgPool operation on B7 to obtain a separated weight concatenated enhancement feature map B8 with a size of 136×100×64; input B8 into the CBS submodule for operation to obtain a separated weight concatenated enhancement feature map B9 with a size of 136×100×64.

[0203] Then, perform a Concat operation on B4, B6, and B9 to obtain a separated weighted concatenated augmented feature map B10 with a size of 272×200×192. Input B10 into the CBS submodule for operation to obtain a separated weighted concatenated augmented feature map B11 with a size of 272×200×192. Apply CBAM to B11 to obtain a separated weighted concatenated augmented feature map B12 with a size of 272×200×192. Input B12 into the CBS submodule for operation to obtain a separated weighted concatenated augmented feature map Y2 with a size of 272×200×64.

[0204] Furthermore, the separated weighted feature map is input into the dual-convolution multi-pooling small target detection module to obtain the feature map for detecting foreign objects attached to utility poles. Y10 is then input into the CBS submodule for processing, resulting in a feature map Z1 of 136×100×128 pixels. Z1 is then processed along two paths:

[0205] (1) In path one, perform the Dilated Conv operation on Z1 to obtain the feature map Z2 with a size of 136×100×128 for detecting foreign objects attached to the utility pole; perform the AvgPool operation on Z2 to obtain the feature map Z3 with a size of 68×50×128 for detecting foreign objects attached to the utility pole; perform the Dynamic Conv operation on Z3 to obtain the feature map Z4 with a size of 68×50×128 for detecting foreign objects attached to the utility pole; perform the AvgPool operation on Z4 again to obtain the feature map Z5 with a size of 34×25×128 for detecting foreign objects attached to the utility pole.

[0206] (2) In path two, the Dynamic Conv operation is performed on Z1 to obtain the pole attachment foreign object detection feature map Z6 with a size of 136×100×128; the MaxPool operation is performed on Z6 to obtain the pole attachment foreign object detection feature map Z7 with a size of 68×50×128; the Dilated Conv operation is performed on Z7 to obtain the pole attachment foreign object detection feature map Z8 with a size of 68×50×128; the MaxPool operation is performed on Z8 again to obtain the pole attachment foreign object detection feature map Z9 with a size of 34×25×128.

[0207] Further, Z5 and Z9 are added element-wise to obtain a feature map Z10 with a size of 34×25×128 for detecting foreign objects attached to utility poles. Z10 is then input into the CBS submodule for processing to obtain a feature map Z11 with a size of 34×25×128. Z11 is then processed using the Sigmoid activation function, and its output is multiplied element-wise with Z11 to obtain a feature map Z12 with a size of 34×25×128. Z12 is then input into the Conv_BN_Mish submodule for processing to obtain a feature map Z13 with a size of 34×25×256. Z13 is then subjected to Dilated Conv and Dynamic Conv operations to obtain feature maps Z14 and Z15 with sizes of 34×25×256, respectively. Perform a concat operation on Z14 and Z15 to obtain a 34×25×512 feature map of foreign objects attached to utility poles, Z16. Input Z16 into the CBS submodule for further processing to obtain a 34×25×512 feature map of foreign objects attached to utility poles, Z17. Input the feature map of foreign objects attached to utility poles into the Head module to obtain the detection results, which include bounding box coordinates and the category of foreign objects attached to the utility poles, such as... Figure 9 As shown.

[0208] For comparison of technical effectiveness, existing technologies can be referenced. Currently, deep learning-driven methods for detecting foreign objects on utility poles are mostly based on improvements to general object detection frameworks. One approach is to directly transfer general models, using models such as YOLOv5 (You Only Look Once version 5) and Faster R-CNN (Faster Region-based Convolutional Neural Network) pre-trained on general datasets like COCO, fine-tuned with a utility pole foreign object annotation dataset for detection. While this approach is simple, it suffers from poor feature adaptability, parameter redundancy, weak scene generalization, and is prone to missing small-sized foreign objects. Inference is slow when deploying edge devices, and detection accuracy fluctuates greatly under different environments.

[0209] Another type is the improvement scheme that introduces basic attention mechanisms or focuses on optimizing small targets. The former incorporates simple attention modules such as CBAM and SE-Net to enhance the attention to foreign objects, but due to the singularity of attention and weak multi-scale collaboration, it is difficult to deal with complexly distributed foreign objects and will increase the complexity of the model. The latter optimizes the detection of small foreign objects through feature pyramid enhancement and small target anchor box design, but is limited by the contradiction between generalization and lightweight design, as well as the weak adaptability to complex backgrounds, resulting in decreased performance across scenes, difficulty in edge deployment, and a high false detection rate.

[0210] Based on the above, existing deep learning-based methods for detecting foreign objects on utility poles have the following problems:

[0211] (1) Poor feature adaptability of multi-scene and multi-scale foreign objects

[0212] Existing methods are not optimized for the complex environment of utility poles (urban-rural differences, changes in lighting) and multi-scale foreign objects (from large billboards to tiny debris), resulting in weak network generalization, an imbalance in the detection of foreign objects of different sizes, and a tendency to make false or false detections.

[0213] (2) It is difficult to balance lightweight design with detection accuracy.

[0214] To improve detection accuracy, existing methods often enhance features by adding attention modules or increasing the number of network layers. However, this significantly increases the number of network parameters and computational cost, resulting in a significant decrease in inference speed. On the other hand, simplifying the network sacrifices its feature extraction capabilities, creating a contradiction between "lightweight for inaccuracy and accurate for inefficiency."

[0215] (3) Insufficient detection capability for small target foreign objects

[0216] In utility pole scenarios, small foreign objects are easily obscured by background interference. Existing methods lack sufficient feature mining capabilities, failing to effectively distinguish foreign objects from similar backgrounds. Furthermore, due to the scarcity of small target samples, network training tends to favor larger targets, resulting in a high false negative rate.

[0217] To address the aforementioned problems, this invention provides a deep learning-based method for detecting foreign objects attached to utility poles. A multi-module synthetic pole foreign object detection network, MSPFOD-Net (Multi-Module Synergy Pole ForeignObject Detection Network), is designed, comprising three key modules: a multi-scale feature perception enhancement module, a lightweight cascaded weighted separable module, and a dual-convolution multi-pooling small target detection module. The multi-scale feature perception enhancement module dynamically fuses features from different levels and receptive fields, expanding the image field of view to capture the distribution of foreign objects in complex backgrounds and strengthening multi-scale feature representation. This enables the network to accurately identify foreign objects of both large and small scales, solving the feature adaptability problem for cross-scene and cross-scale foreign object detection. The lightweight cascaded weighted separable module reduces the parameter size through depthwise separable convolution and attention collaboration. While reducing computational and parameter requirements, it strengthens feature interaction through a cascaded structure, ensuring detection accuracy and resolving the contradiction between lightweight and high-precision detection, meeting the needs of real-time detection. Dual-convolution multi-pooling small target detection module: This module combines dilated convolution (to expand the effective field of view) and dynamic convolution (to adaptive feature extraction) with a multi-pooling strategy to enhance the saliency of small target features; at the same time, it uses feature modulation and fusion mechanisms to suppress background interference, improve the detection efficiency and accuracy of small target foreign objects, and solve the problem of missed detection and false detection of small targets in complex backgrounds.

[0218] Specifically, please refer to Figure 10 This invention utilizes a multi-scale feature perception enhancement module to integrate convolutional features from different receptive fields through hierarchical feature fusion, expanding the perception range and accurately capturing the distribution of foreign objects in complex backgrounds. It strengthens multi-scale feature interaction, solving the feature adaptation problem for cross-scene foreign object detection, and enabling relatively stable identification of foreign objects of various sizes. By employing a lightweight, weighted, cascaded module with depthwise separable convolution and attention mechanisms, it reduces computational load while enhancing feature interaction through a cascaded structure, balancing lightweight design with detection accuracy. This meets the real-time requirements of power line inspection and provides a solution for edge device deployment. A dual-convolution multi-pooling small target detection module fuses dilated convolution and dynamic convolution to expand the field of view and adaptively extract features. Combined with multi-pooling to enhance the saliency of small targets and suppress background interference, it improves the detection efficiency and accuracy of small foreign objects, solving the problem of missed and false detections of foreign objects in complex scenes.

[0219] In summary, the multi-scale feature perception enhancement module proposed in this invention addresses the significant scale differences of foreign objects in utility pole scenarios by constructing a dynamic feature fusion mechanism to intelligently integrate convolutional features from different levels and receptive fields. On one hand, by expanding the image perception range, it accurately captures the spatial distribution of foreign objects such as hanging objects and tangled cables in the complex background of utility poles. On the other hand, it strengthens multi-scale feature representation, enabling the complete extraction of the overall outline of large bird nests as well as the keen capture of local details of small foreign objects such as tangled thin lines. In this way, it effectively solves the feature adaptation problem of cross-scene and cross-scale foreign object detection in utility pole inspection, ensuring the network's accurate identification of foreign objects of different sizes and shapes. The designed lightweight module with cascaded weights adopts a collaborative design of depthwise separable convolution and attention mechanisms. While significantly reducing the amount of computation and parameters, it strengthens feature interaction through the cascaded structure, ensuring accurate detection of various foreign objects such as floating plastic bags and tangled kite strings on utility poles. This design cleverly resolves the contradiction between lightweight design and high-precision detection, enabling the network to operate efficiently on edge devices such as drones and intelligent inspection terminals, achieving rapid detection and response to foreign objects on utility poles during power inspections. The proposed dual-convolution multi-pooling small object detection module innovatively integrates dilated convolution and dynamic convolution techniques. Dilated convolution broadens the field of view, capturing the spatial relationship between small objects and their surrounding environment; dynamic convolution adaptively adjusts the feature extraction strategy, accurately adapting to changes in the shape of small foreign objects. Combined with a multi-pooling strategy to enhance the saliency of small objects, and utilizing feature modulation and fusion mechanisms to suppress interference from complex backgrounds, this significantly improves the efficiency and accuracy of small object detection, effectively solving the problem of traditional methods easily missing or falsely detecting small foreign objects in complex backgrounds. Therefore, MSPFOD-Net constructs a three-level pipelined feature processing architecture, achieving end-to-end deep optimization from raw image acquisition to foreign object detection result output. The multi-scale feature perception module uses a hierarchical fusion strategy to comprehensively capture and enhance the multi-dimensional features of foreign objects on utility poles; the separable weight cascade module uses depthwise separable convolution and attention mechanisms in synergy to enhance feature semantic expression while reducing computational resource consumption; and the dual-convolution multi-pooling module, with its innovative combination of dilated convolution and dynamic convolution, accurately locates small foreign object targets.

[0220] In this embodiment of the invention, the MSPFOD-Net proposed in this invention constructs core advantages through a hierarchical architecture and differentiated processing mechanism: the multi-scale feature perception enhancement module, with its dual-path and attention mechanism, can accurately capture the texture details of tiny foreign objects such as tangled thin lines and fluttering plastic bags, as well as the macroscopic morphological features of large bird nests and hanging objects, effectively filtering out complex background interference; the separation weight cascaded lightweight module balances performance and computational cost with a dual-track architecture, which can both mine the implicit correlation between foreign object features through deep separation units and quickly transmit basic features with lightweight paths, meeting the real-time detection needs of edge devices such as drones and smart terminals for utility pole inspection; the dual-convolution multi-pooling detection module integrates multi-pooling and deformable convolution, taking into account both the global distribution and local salient features of foreign objects, dynamically adapting to the irregular deformation of foreign objects caused by wind and environmental changes, and accurately locating tiny or deformed attached foreign objects. Each module addresses the pain points of detecting foreign objects attached to utility poles, which are characterized by "small targets, varied shapes, and complex backgrounds." It forms a pipeline-style collaborative architecture, achieving a closed loop from feature extraction and optimization to accurate detection. This significantly improves the detection accuracy and efficiency of various types of foreign objects attached to utility poles, providing reliable technical support for intelligent power inspection and effectively ensuring the safe and stable operation of the power grid.

[0221] Please see Figure 11 , Figure 11 This is a structural block diagram of a foreign object detection device for utility poles based on deep learning, provided in Embodiment 2 of the present invention.

[0222] This invention provides a deep learning-based foreign object detection device for utility poles, comprising:

[0223] The acquisition module 1101 is used to acquire images of foreign objects attached to utility poles and input the images of foreign objects attached to utility poles into the detection network of foreign objects attached to utility poles. The detection network of foreign objects attached to utility poles includes a multi-scale feature perception enhancement module, a separation weight cascade lightweight module, a dual-convolution multi-pooling small target detection module, and a target detection head.

[0224] Module 1102 is used to extract features from images of foreign objects attached to utility poles using a multi-scale feature perception enhancement module, and generate a multi-scale feature perception enhancement map.

[0225] The input module 1103 is used to take the multi-scale feature perception enhancement map as the input of the separation weight cascade lightweight module and output the separation weight cascade enhanced feature map.

[0226] The first detection module 1104 is used to perform dual-convolution multi-pooling small target detection on the separated weighted cascaded enhanced feature map through the dual-convolution multi-pooling small target detection module, and generate a feature map for detecting foreign objects attached to utility poles.

[0227] The second detection module 1105 is used to perform target detection on the feature map of foreign objects attached to the utility pole using a target detection head, and output the detection results of foreign objects attached to the utility pole.

[0228] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0229] This invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program; when the computer program is executed by the processor, the processor performs the steps of the deep learning-based method for detecting foreign objects attached to utility poles as described in any of the above embodiments.

[0230] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the steps of the deep learning-based method for detecting foreign objects attached to utility poles as described in any of the above embodiments.

[0231] This invention also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the deep learning-based method for detecting foreign objects attached to utility poles as described in any of the above embodiments.

[0232] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting foreign matter attached to a utility pole based on deep learning, characterized by, The utility model relates to: acquire the image of pole attached foreign matter, and input pole attached foreign matter image to pole attached foreign matter detection network, and pole attached foreign matter detection network includes multiscale feature perception enhancement module, separate weight cascade light module, double roll multi pool small target detection module and target detection head; the multiscale feature perception enhancement module is used to feature extraction of pole attached foreign matter image, and the multiscale feature perception enhancement module generates multiscale feature perception enhancement graph; the multiscale feature perception enhancement graph is used as the input of the separate weight cascade light module, and the separate weight cascade enhancement feature graph is output; the double roll multi pool small target detection module is used to double roll multi pool small target detection of the separate weight cascade enhancement feature graph, and pole attached foreign matter detection feature graph is generated; the target detection head is used to target detection of the pole attached foreign matter detection feature graph, and the pole attached foreign matter detection result is output.

2. The deep learning-based utility pole attachment foreign matter detection method according to claim 1, characterized by, The multiscale feature perception enhancement module includes 1x1 convolution layer, convolution submodule, basic convolution block, dynamic deformable feature fusion submodule, sigmoid activation function layer; the multiscale feature perception enhancement module is used to feature extraction of pole attached foreign matter image, and the multiscale feature perception enhancement module generates multiscale feature perception enhancement graph, including: 1x1 convolution layer is used to the convolution operation of pole attached foreign matter image, and the first pole convolution feature graph is output, and the first pole convolution feature graph is respectively used as the input of convolution submodule and basic convolution block, and the second pole convolution feature graph and the third pole convolution feature graph are output; the second pole dynamic deformable feature graph is output through the dynamic deformable feature fusion submodule to the second pole convolution feature graph; the second pole dynamic deformable feature graph is output through the dynamic deformable feature fusion submodule to the third pole convolution feature graph; the fourth pole convolution feature graph is output by the first pole dynamic deformable feature graph as the input of basic convolution block; the fifth pole convolution feature graph is output by the second pole dynamic deformable feature graph as the input of convolution submodule; the fifth pole convolution feature graph and the fourth pole convolution feature graph are added element by element, and the pole convolution splicing feature graph is output; 1x1 convolution layer is used to the convolution operation of pole convolution splicing feature graph, and the sixth pole convolution feature graph is output, and the sixth pole convolution feature graph is used as the input of sigmoid activation function layer, and the pole transformation feature graph is output; the pole multiplication feature graph is output by the pole transformation feature graph and the sixth pole convolution feature graph element by element, and the pole multiplication feature graph is convolved through 1x1 convolution layer, and the multiscale feature perception enhancement graph is generated.

3. The deep learning-based utility pole attachment foreign matter detection method according to claim 2, characterized by, The dynamic deformable feature fusion sub-module includes a 3*3 convolution layer, a 7*7 convolution layer, and a basic convolution block; the second pole convolution feature map is subjected to dynamic deformable feature fusion through the dynamic deformable feature fusion sub-module, and a first pole dynamic deformable feature map is output, including: The second pole convolution feature map is taken as input of the basic convolution block, a first pole basic convolution feature map is output, and the first pole basic convolution feature map is input into the 3*3 convolution layer and the 7*7 convolution layer respectively for convolution operation, a seventh pole convolution feature map and an eighth pole convolution feature map are generated; The first pole basic convolution feature map is subjected to deformable convolution operation, a pole deformable convolution feature map is output, and the pole deformable convolution feature map is taken as input of the basic convolution block, a ninth pole convolution feature map is output; The seventh pole convolution feature map is subjected to down-sampling, a pole down-sampling feature map is output, and the eighth pole convolution feature map is subjected to up-sampling, a pole up-sampling feature map is generated; The pole down-sampling feature map and the ninth pole convolution feature map are spliced, a first pole splicing feature map is output, and the pole up-sampling feature map and the ninth pole convolution feature map are spliced, a second pole splicing feature map is output; The first pole splicing feature map and the second pole splicing feature map are taken as input of the basic convolution block respectively, a first intermediate pole convolution feature map and a second intermediate pole convolution feature map are output; The first intermediate pole convolution feature map and the second intermediate pole convolution feature map are subjected to element-wise addition, an intermediate pole addition feature map is output, and the intermediate pole addition feature map is subjected to dynamic convolution operation, a pole dynamic convolution feature map is generated; The pole dynamic convolution feature map is taken as input of the basic convolution block, a first pole dynamic deformable feature map is output.

4. The deep learning-based utility pole attachment foreign matter detection method according to claim 1, characterized by, The separated weight cascade lightweight module includes a 1*1 convolution layer, a depth separable attention unit, a convolution submodule, and a basic convolution block; the multi-scale feature perception enhancement map is taken as input of the separated weight cascade lightweight module, a separated weight cascade enhancement feature map is output, including: The multi-scale feature perception enhancement map is subjected to convolution operation through the 1*1 convolution layer, a first separated weight cascade initial feature map is generated, and the first separated weight cascade initial feature map is subjected to feature strengthening extraction through the depth separable attention unit, a second separated weight cascade initial feature map is generated; The second separated weight cascade initial feature map is taken as input of the convolution submodule, a third separated weight cascade initial feature map is output, and the third separated weight cascade initial feature map is subjected to feature strengthening extraction through the depth separable attention unit, a fourth separated weight cascade initial feature map is generated; The fourth separated weight cascade initial feature map is taken as input of the basic convolution block, a first separated weight cascade feature map is output, and the first separated weight cascade feature map is subjected to feature strengthening extraction through the depth separable attention unit, a second separated weight cascade feature map is generated. The fourth separated weight concatenated initial feature map is taken as an input of a basic convolution block, and a fifth separated weight concatenated initial feature map is outputted, and the fifth separated weight concatenated initial feature map is subjected to feature enhancement extraction through the deep separable attention unit, and a sixth separated weight concatenated initial feature map is generated. The first separated weight concatenated initial feature map is taken as an input of a convolution sub-module, and a seventh separated weight concatenated initial feature map is outputted. The seventh separated weight concatenated initial feature map is taken as an input of a basic convolution block, and an eighth separated weight concatenated initial feature map is outputted. The eighth separated weight concatenated initial feature map and the sixth separated weight concatenated initial feature map are spliced, and a ninth separated weight concatenated initial feature map is outputted, and the ninth separated weight concatenated initial feature map is subjected to convolution operation through a 1*1 convolution layer, and a separated weight concatenated enhanced feature map is generated.

5. The deep learning-based utility pole attachment foreign matter detection method according to claim 4, characterized by, The deep separable attention unit comprises a basic convolution block, a deep separable convolution layer and a spatial-channel attention mechanism layer; the first separated weight concatenated initial feature map is subjected to feature enhancement extraction through the deep separable attention unit, and a second separated weight concatenated initial feature map is generated, which comprises: The first separated weight concatenated initial feature map is taken as an input of a basic convolution block, and an intermediate separated weight concatenated initial feature map is outputted. The intermediate separated weight concatenated initial feature map is subjected to deep convolution through a deep separable convolution layer, and a first deep separable attention feature map is generated, and the first deep separable attention feature map is subjected to maximum pooling and average pooling respectively, and a second deep separable attention feature map and a third deep separable attention feature map are outputted. The intermediate separated weight concatenated initial feature map is subjected to attention calculation through a spatial-channel attention mechanism layer, and a fourth deep separable attention feature map is outputted, and the fourth deep separable attention feature map is taken as an input of a basic convolution block, and a fifth deep separable attention feature map is outputted. The second deep separable attention feature map and the third deep separable attention feature map are taken as inputs of a basic convolution block respectively, and a sixth deep separable attention feature map and a seventh deep separable attention feature map are outputted. The sixth deep separable attention feature map, the seventh deep separable attention feature map and the fifth deep separable attention feature map are spliced, and an eighth deep separable attention feature map is outputted. The eighth deep separable attention feature map is taken as an input of a basic convolution block, and a ninth deep separable attention feature map is generated, and the ninth deep separable attention feature map is subjected to attention calculation through a spatial-channel attention mechanism layer, and a tenth deep separable attention feature map is generated. The tenth deep separable attention feature map is taken as an input of a basic convolution block, and a second separated weight concatenated initial feature map is outputted.

6. The deep learning-based utility pole attachment foreign matter detection method according to claim 1, characterized by, The double-convolution multi-pooling small target detection module comprises a basic convolution block, a convolution submodule, and a Sigmoid activation function layer. The separated weight cascade enhanced feature map is taken as the input of the basic convolution block, a first double-convolution multi-pooling small target detection feature map is output, and dilated convolution is performed on the first double-convolution multi-pooling small target detection feature map to generate a second double-convolution multi-pooling small target detection feature map; The second double-convolution multi-pooling small target detection feature map is subjected to average pooling to output a third double-convolution multi-pooling small target detection feature map, and dynamic convolution is performed on the third double-convolution multi-pooling small target detection feature map to generate a fourth double-convolution multi-pooling small target detection feature map; The fourth double-convolution multi-pooling small target detection feature map is subjected to average pooling to output a fifth double-convolution multi-pooling small target detection feature map; The first double-convolution multi-pooling small target detection feature map is subjected to dynamic convolution to generate a sixth double-convolution multi-pooling small target detection feature map, and maximum pooling is performed on the sixth double-convolution multi-pooling small target detection feature map to output a seventh double-convolution multi-pooling small target detection feature map; The seventh double-convolution multi-pooling small target detection feature map is subjected to dilated convolution to generate an eighth double-convolution multi-pooling small target detection feature map, and maximum pooling is performed on the eighth double-convolution multi-pooling small target detection feature map to output a ninth double-convolution multi-pooling small target detection feature map; The ninth double-convolution multi-pooling small target detection feature map and the fifth double-convolution multi-pooling small target detection feature map are subjected to element-wise addition to generate a tenth double-convolution multi-pooling small target detection feature map, and the tenth double-convolution multi-pooling small target detection feature map is taken as the input of the basic convolution block to output an eleventh double-convolution multi-pooling small target detection feature map; The eleventh double-convolution multi-pooling small target detection feature map is taken as the input of the Sigmoid activation function layer to output an intermediate double-convolution multi-pooling small target detection feature map, and the intermediate double-convolution multi-pooling small target detection feature map and the tenth double-convolution multi-pooling small target detection feature map are subjected to element-wise multiplication to output a twelfth double-convolution multi-pooling small target detection feature map; The twelfth double-convolution multi-pooling small target detection feature map is taken as the input of the convolution submodule to output a thirteenth double-convolution multi-pooling small target detection feature map, and the thirteenth double-convolution multi-pooling small target detection feature map is subjected to dilated convolution and dynamic convolution, respectively, to generate a fourteenth double-convolution multi-pooling small target detection feature map and a fifteenth double-convolution multi-pooling small target detection feature map; The fourteenth double-convolution multi-pooling small target detection feature map and the fifteenth double-convolution multi-pooling small target detection feature map are spliced to output a sixteenth double-convolution multi-pooling small target detection feature map; The sixteenth double-convolution multi-pooling small target detection feature map is taken as the input of the basic convolution block to output a power line pole attached foreign matter detection feature map.

7. A deep learning-based device for detecting foreign substances attached to a utility pole, characterized by It comprises: The acquisition module is used to acquire a pole-attached foreign matter image and input the pole-attached foreign matter image into a pole-attached foreign matter detection network, wherein the pole-attached foreign matter detection network comprises a multi-scale feature perception enhancement module, a separated weight cascade lightweight module, a double-rolled multi-pooling small target detection module and a target detection head. The multi-scale feature perception enhancement module is used to perform feature extraction on the pole-attached foreign matter image and generate a multi-scale feature perception enhancement image. The input module is used to input the multi-scale feature perception enhancement image as an input of the separated weight cascade lightweight module and output a separated weight cascade enhancement feature image. The first detection module is used to perform double-rolled multi-pooling small target detection on the separated weight cascade enhancement feature image by the double-rolled multi-pooling small target detection module and generate a pole-attached foreign matter detection feature image. The second detection module is used to perform target detection on the pole-attached foreign matter detection feature image by the target detection head and output a pole-attached foreign matter detection result.

8. A computer device, comprising: The computer program is executed to implement the pole-attached foreign matter detection method based on deep learning.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program product comprises a computer program stored on a non-transitory computer-readable storage medium, and the computer program comprises program instructions, wherein when the program instructions are executed by a computer, the computer executes the pole-attached foreign matter detection method based on deep learning.

10. A computer program product, characterised in that, The computer program product comprises a computer program stored on a non-transitory computer-readable storage medium, and the computer program comprises program instructions, wherein when the program instructions are executed by a computer, the computer executes the pole-attached foreign matter detection method based on deep learning.

Images