Method for identifying and counting cables in cable trench based on improved YOLO model

Abstract

The invention relates to a method for identifying and counting cables in a cable trench based on an improved YOLO model, and belongs to the technical field of power cables. The method comprises the following steps: firstly, acquiring image information in a cable trench by using a camera, and then inputting the image information into an improved YOLOv8 model for cable identification. The improved YOLOv8 model introduces an EFFVM module into a backbone network part, so that the recognition capability of the model on a shielded target is enhanced; an FGM module is introduced into a neck network part, so that information exchange of deep and shallow layer features of the neck of the model is enhanced; and a prediction output head for the angle is newly added in a head network part, and a ProbIoU loss function is used to predict the rotating rectangular frame. And finally, counting and outputting the number of cables according to a cable identification result output by the model. According to the method, the number of the cables in the cable trench can be quickly and accurately identified, and technical support is provided for operation and maintenance of the power cables.

Description

Technical Field

[0001] The invention belongs to the technical field of power cables and relates to a method for identifying and counting cables in a cable trench based on an improved YOLO model. Background Art

[0002] Image recognition technology has experienced significant development in recent years, mainly due to the progress of deep learning. The introduction of convolutional neural networks (CNN) marked an important milestone, and from AlexNet to VGG, GoogLeNet and ResNet, these networks achieved unprecedented results in the ImageNet competition, allowing computers to perform better than ever before in image classification tasks.

[0003] With the improvement of hardware performance, especially the popularity of GPUs, training deep neural networks has become more feasible. The establishment of large-scale data sets (such as ImageNet) also provides rich data resources for model training. These data sets help researchers develop and optimize more complex network structures. In the field of target detection, the development of the YOLO (You Only Look Once) recognition algorithm is particularly prominent. YOLO was proposed by Joseph Redmon and others. Its core concept is to regard the target detection task as a regression problem, thereby significantly improving the detection speed. From YOLOv1 to YOLOv8, the YOLO series of algorithms have gradually solved the defects in early versions through continuous iteration and optimization, making it perform better in target detection tasks. These algorithms have not only achieved remarkable results in academia, but have also been widely used in industry.

[0004] Overall, image recognition technology has made significant progress in accuracy, speed, and breadth of application. From medical image analysis to autonomous driving, from intelligent monitoring to augmented reality, the application scenarios of image recognition technology are becoming more and more extensive, promoting innovation and development in various industries. In the future, with the emergence of more advanced algorithms and more powerful computing resources, image recognition technology is expected to continue to achieve new breakthroughs.

[0005] There are three main types of image recognition algorithms currently in common use:

[0006] 1) Convolutional Neural Network (CNN). The convolutional neural network method mainly extracts local features of the image through convolution operations in the convolution layer. The pooling layer reduces the amount of data and computational complexity through downsampling operations while retaining important features. The fully connected layer classifies the extracted features and outputs the final classification results. Representative models include AlexNet, VGG, ResNet, etc.

[0007] 2) YOLO (You Only Look Once). This method divides the entire image into multiple grids in the single-stage detection phase. Each grid is responsible for predicting multiple bounding boxes and their category probabilities. It then directly regresses and predicts the location and category of the bounding boxes, rather than generating candidate regions and then classifying them. The latest YOLO model has been iterated to YOLOv8.

[0008] 3) Image Generative Adversarial Networks (GANs). This method uses a generator to generate realistic images, then uses a discriminator to determine whether the image is real or generated, and finally conducts adversarial training on the generator and the discriminator. Through adversarial training, the generator continuously improves the quality of generated images, and the discriminator continuously improves its discrimination ability.

[0009] At present, there are still the following problems and difficulties in image recognition algorithms:

[0010] Regarding image recognition algorithms, although an existing cable identification and counting method in a cable trench based on an improved YOLO model can provide relatively accurate predictions, it still faces a series of problems and difficulties in practical applications.

[0011] Image recognition models are too data-dependent. Modern image recognition algorithms, especially deep learning models, require a large amount of labeled data for training. It is costly to obtain and annotate large-scale, high-quality datasets. In addition, biases in the training datasets may cause the models to perform poorly in practical applications. For example, certain categories may be under- or over-represented in the training dataset. Image recognition technology also has the problem of insufficient explainability and interpretability. Deep learning models are often considered "black boxes" and it is difficult to explain their internal decision-making process. This is particularly problematic in areas that require high transparency and reliability, such as medical diagnosis. Image recognition models tend to fail when faced with adversarial samples or slight image perturbations, indicating that the model's decision-making process lacks robustness and interpretability.

[0012] Not only that, real-time and latency issues are one of the problems of this technology. Although some algorithms (such as YOLO) have optimized real-time processing capabilities, real-time processing is still a challenge in high-resolution images or complex scenes. In some delay-sensitive application scenarios (such as autonomous driving), even millisecond delays may pose safety risks. Limited by existing technologies, many models generally have problems with insufficient versatility and generalization capabilities. Although transfer learning can help apply existing models in new fields, the generalization ability of models between different fields is still limited. When the distribution of training and test data is different, the model performance will drop significantly. How to improve the domain adaptability of the model is an important research direction.

[0013] In order to provide technical support for operation and maintenance, image recognition technology is used to identify the number of cables in complex cable trenches due to the complex conditions of cable trenches in the power industry. The YOLO algorithm is used to quickly provide information to operation and maintenance personnel using the high real-time processing capabilities of the YOLOv8 model; the model only needs to quickly and accurately identify cables and their quantities, so it does not involve universal issues; the targeted cable recognition model has excellent recognition capabilities for obscured targets and cables after being improved. Summary of the invention

[0014] In view of this, the object of the present invention is to provide a method for identifying and counting cables in a cable trench based on an improved YOLO model.

[0015] In order to achieve the above object, the present invention provides the following technical solutions:

[0016] A method for identifying and counting cables in a cable trench based on an improved YOLO model, the method comprising the following steps:

[0017] Use the camera to collect image information in the cable trench;

[0018] The collected image information is input into the improved YOLOv8 model for cable identification. The improved YOLOv8 model includes:

[0019] The backbone network includes multiple low-level feature maps and high-level feature maps fusion (CSPDarknet53 to2-Stage FPN, C2f) and characteristic receptive field fusion module (EFFVM), where the EFFVM module includes multiple convolution modules (Convoluton, CONv) with different expansion coefficients;

[0020] The neck network part includes Feature Pyramid Networks (FPN) and Path Aggregation Network (PAN). The Concat connection in the FPN structure is replaced by the Fast Gradient Method (FGM) module. The FGM module includes multiple Enhanced guidance attention mechanism (EGAM).

[0021] The head network part includes three detection heads, each of which includes a classification branch and a regression branch, and a new angle prediction output head is added;

[0022] According to the cable identification results output by the improved YOLOv8 model, the number of cables is counted and output.

[0023] Furthermore, the structure of the EFFVM module is as follows:

[0024] Enter feature information;

[0025] Feature information is processed through a CBS module;

[0026] Connect three CONv modules with different expansion coefficients in series;

[0027] Use Concat to concatenate the features processed by the three CONv modules with the input features;

[0028] After a CONv module, feature fusion processing is performed;

[0029] Output the fused feature information.

[0030] Further, the structure of the FGM module is as follows:

[0031] Enter feature information;

[0032] Use the same bidirectional EGAM module to provide feature guidance for each layer.

[0033] Add and fuse the guidance feature information and the own feature information;

[0034] Output the fused feature information.

[0035] Furthermore, the structure of the dynamic snake convolution DSConv is as follows:

[0036] Enter feature information;

[0037] Use deformation offset △ to deform the convolution kernel;

[0038] Perform convolution operation;

[0039] Output the feature information after convolution.

[0040] Further, the structure of the SEAM module is as follows:

[0041] Enter feature information;

[0042] Feature extraction is performed through CSMM channels with residual connections and spatial mixing modules;

[0043] The information of each channel is fused through a two-layer fully connected network;

[0044] The output of the SEAM module is used as attention and multiplied with the original features;

[0045] Output the feature information after attention enhancement.

[0046] Furthermore, the improved YOLOv8 model uses the ProbIoU loss function to predict the rotated rectangular box;

[0047] ProbIoU converts the real box and the predicted box into a Gaussian distribution ProbIoU designs a network that regresses three parameters (a′, b′, θ) to make the shape and rotation angle of the rotation prediction box closer to the real box; the similarity of the two Gaussian distributions is obtained by calculating the bus coefficient;

[0048] The Hellinger distance is introduced to bring the two Gaussian distributions closer, so that the real box and the predicted box overlap more. The calculation formula of the Hellinger distance is as shown in formula (1), the calculation formula of ProbIoU is as shown in formula (2), and the final loss function formula is as shown in formula (3):

[0049]

[0050] ProbIoU=1-H D (p,q) (2)

[0051] L ProbIoU (p,q)=1-ProbIoU(p,q) (3)

[0052] Where p(x) and q(x) represent the probability density functions of the Gaussian distribution of the real box and the predicted box, respectively.

[0053] Furthermore, the improved YOLOv8 model uses dynamic snake convolution DSConv to extract local features of cable images.

[0054] Furthermore, the improved YOLOv8 model uses the SEAM attention module to enhance the focus on the target area in the image and suppress the background area.

[0055] Furthermore, the improved YOLOv8 model uses the EFFVM module to enhance the model's ability to recognize occluded targets.

[0056] The beneficial effect of the present invention is that the present invention is based on the YOLOv8 model and is targeted for improvement. On the basis of retaining high real-time performance as much as possible, the model's recognition ability for occluded targets is improved by replacing the EFFVM module in the backbone part and adding the SEAM attention module in the head, so that the model can better adapt to the complex cable laying environment in the cable trench. In order to make the model more capable of capturing local features, this paper adopts the FGM module so that each detection head can obtain sufficient exchange information, improve the model recognition accuracy, and the model's fast and accurate recognition can provide stable cable quantity information to the operation and maintenance personnel.

[0057] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:

[0059] Figure 1 This is the YOLOv8 structure diagram;

[0060] Figure 2 Comparison diagram of the horizontal detection frame and the rotation detection frame; (a) is the horizontal frame; (b) is the rotation frame;

[0061] Figure 3 To improve the YOLOv8 network architecture;

[0062] Figure 4 It is the EFFVM module;

[0063] Figure 5 It is a snake-shaped dynamic convolution;

[0064] Figure 6 It is the SEAM module. DETAILED DESCRIPTION

[0065] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and the features in the embodiments can be combined with each other without conflict.

[0066] Among them, the drawings are only used for illustrative explanations, and they only represent schematic diagrams rather than actual pictures, and should not be understood as limitations on the present invention. In order to better illustrate the embodiments of the present invention, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0067] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if the terms "upper", "lower", "left", "right", "front", "rear", etc. indicate the orientation or position relationship, they are based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation. Therefore, the terms describing the position relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0068] The whole program is divided into the following two parts:

[0069] Introduction to the overall framework of the YOLOv8 model;

[0070] An introduction to the overall architecture of the improved YOLOv8 model and the advantages of its added and modified modules;

[0071] 1.YOLOv8 detection model

[0072] The YOLOv8 network model is simply divided into three parts: backbone, neck and head. In the backbone part of YOLOv8, the CBS module, C2f module and SPPF module are mainly used. The CBS module is used for downsampling processing, which consists of a 3×3 Conv convolution module, a BN normalization layer and a SiLU activation function. The C2f module extracts features by multi-layering residual modules after gradient splitting (Split), and finally fuses the branches to obtain enhanced features. The SPPF module undergoes three 5×5 pooling downsampling processes in turn, and integrates multi-scale feature information through the CBS module.

[0073] The neck part is composed of FPN and PAN structures. FPN and PAN structures are bottom-up and top-down feature layer fusion methods respectively. After Concat connection with feature layers of different scales in the backbone part, the C2f module is used for feature extraction, so that the information of deep and shallow features is fused with each other.

[0074] The head part has three detection heads, each of which has two branches: classification branch and regression branch. This structure adopts a dual-branch decoupling head to reduce the weight of the head. It uses two parallel branches to perform classification and regression, which improves the detection effect of the model. The YOLOv8 structure is shown in the figure below. Figure 1 shown.

[0075] Cables are the main identification targets during cable laying. The traditional YOLO model uses a horizontal detection frame. In this case, it may cause misjudgment or omission of identification by operation and maintenance personnel and generate a large amount of redundant information.

[0076] Compared with horizontal box target detection, rotation target detection mainly targets a large number of densely packed small target objects and some objects with a large aspect ratio. By increasing the regression angle, the prediction box can better select the target, achieve a more compact bounding box, and reduce the redundant information of the selected object. The comparison between the traditional horizontal box detection algorithm and the rotation target detection algorithm is shown in the figure below. Figure 2 As shown, (a) is the horizontal box HBB; (b) is the rotation box OBB.

[0077] The YOLOv8-obb ​​algorithm modifies the detection head based on YOLOv8, and adds a new angle prediction output head in the decoupling head. It also replaces the original CIoU loss function and uses the ProbIoU (Probabilistic Intersection over Union) loss to predict the rotated rectangular box to achieve the purpose of directional detection.

[0078] ProbIoU first converts the real box and the predicted box into a Gaussian distribution ProbIoU designs a network that regresses three parameters (a′, b′, θ) to make the shape and rotation angle of the rotation prediction box closer to the real box. The similarity of two Gaussian distributions is obtained by calculating the Bhattacharyya Coefficient.

[0079] Finally, the Hellinger distance is introduced to bring the two Gaussian distributions closer together, so that the real box and the predicted box overlap more. The calculation formula of the Hellinger distance is as follows: (1), the calculation formula of ProbIoU is as follows: (2), and the final loss function formula is as follows:

[0080]

[0081] ProbIoU=1-H D (p,q) (2)

[0082] L ProbIoU(p,q)=1-ProbIoU(p,q) (3)

[0083] Where p(x) and q(x) represent the probability density functions of the Gaussian distribution of the real box and the predicted box, respectively.

[0084] 2. Improved YOLOv8 detection model

[0085] In view of the complex situation and identification objects in the cable trench, some modules of the YOLO framework are improved on the basis of introducing the above-mentioned YOLOv8-obb ​​algorithm. First, this paper proposes the EFFVM module and introduces it into the C2f module of the backbone part. By fusing the receptive fields of different scales, the feature extraction ability of the model is enhanced. Then, this paper proposes the FGM module to replace the Concat connection in the FPN structure, and strengthens the information exchange of the deep and shallow features of the neck of the model. Secondly, the snake-shaped dynamic convolution (DSConv) is introduced to expand the receptive field of the model, enhance the perception of key features, adaptively adjust the shape of the convolution kernel, more accurately capture the local features and structural information of the image, and humanely process complex images, so as to obtain better feature extraction ability, robustness and recognition ability for slender objects.

[0086] Considering that when identifying cables in a cable trench environment, there may often be target occlusion, resulting in local aliasing and feature loss. Therefore, this paper introduces the SEAM module to enhance the focus on the target area in the image through multi-scale detection, while suppressing the background area. In combination with the EFFVM module mentioned above, the model's ability to capture local features and judge global features from local features is specifically improved, thereby improving the model's ability to identify target occlusion.

[0087] After using the camera and the model to identify the cables in the cable trench, the number of cables can be counted and output. Figure 3 shown.

[0088] 2.1 EFFVM module

[0089] In the complex environment of the cable trench, it is necessary to consider the problem of mutual occlusion between cables, so the model's detection ability for target occlusion needs to be especially improved. In the backbone part of the YOLOv8 model, since the depth information obtained after multiple downsampling processes will ignore the features of shallow information, this article considers the EFFVM module, which first processes the feature information through a CBS module to remove the current redundant channel information. Next, three CONv modules with different expansion coefficients are connected in series, and Concat is used to concatenate the features processed by the three CONv modules and the input features. Finally, a CONv module is used to perform feature fusion processing to obtain features with a larger field of view and higher quality. The structure diagram of the EFFVM module is shown below. Figure 4 shown.

[0090] 2.2FGM module

[0091] In order to enable the model to have a higher recognition rate for the local features and overall features of the target object, it is necessary to fully integrate the feature information of the deep and shallow layers so that the deep and shallow features can exchange information and all detection heads can obtain more accurate cross-information. This paper proposes the FGM module, which uses the bidirectional and identical EGAM module (Enhanced guidance attention mechanism) to provide feature guidance for another layer respectively, so that the feature information of another layer can be fully obtained, and the feature itself can be corrected and improved without being affected by other factors.

[0092] The EGAM module first processes the number of channels to the same number as the other layer through Convk=1, and then divides it into two branches: the first branch is used to process the guidance information, and obtains the maximum feature and mean feature information through two parallel maximum pooling downsampling (Maxpooling) and average pooling downsampling (Avgpooling) respectively. Then these two features are fused and added, and the Conv k = 1 operation to reduce the dimension of features and the ReLU function to improve the model's ability to handle nonlinear features. k = 1 operation to increase the dimension of features, and finally Reshape to restore the feature size. The second branch is used to process the feature information itself. k =1 dimensionality reduction, ReLU nonlinear activation function and Conv k =1 The dimension-raising operation is used to obtain the processed self-feature information. Finally, the self-feature information and the guidance feature information are added and fused and the EGAM is obtained through the Sigmod function. x Guidance features.

[0093] 2.3 Dynamic Snake Convolution (DSConv)

[0094] Dynamic snake convolution (DSConv) is used to extract local features of tubular structures to provide greater focusing flexibility for the convolution kernel. The deformation offset △ is introduced, and the deformation offset can be freely learned under thin tubular structures. The concept field often shifts the target, so an iterative strategy is used to observe the targets to be processed in turn, thereby ensuring the continuity of attention without distracting attention from large to small deformation offsets.

[0095] Considering the complex geometric features of the cable image in the cable trench, deformation offset is introduced. The model randomly learns the deformation offset, and the perception area may deviate from the target. This paper adopts Figure 5 The iteration strategy shown.

[0096] This strategy matches each target with an observable position in sequence, ensuring continuous feature attention without over-dispersing the perception area due to large deformation offsets. DSConv is a two-dimensional deformation process covering 9×9. Its advantage is that it is applicable to Xichang tubular structures based on dynamic mechanics and better perceives key features, such as vehicle detection that includes long distances, small targets, and slender targets, such as cables. DSConv covers a 9*9 range during the deformation process, expands the model's receptive field, enhances the perception of key features, and lays the foundation for accurate recognition. The introduction of dynamic snake convolution can adaptively adjust the shape of the convolution kernel, more accurately capture local image features and structural information, and humanely process complex images, thereby obtaining better feature extraction capabilities and robustness.

[0097] 2.4 SEAM Attention Module

[0098] Similar to the EFFVM module, the SEAM attention module is also designed to improve the model's ability to identify target occlusions. The SEAM module of the YOLO-FaceV2 algorithm introduced in this paper enhances attention to the target area in the image through multi-scale detection, while suppressing the background area, thereby improving the model's ability to detect occluded targets.

[0099] The SEAM module structure is as follows Figure 6 shown.

[0100] The first part of SEAM is a CSMM channel and spatial mixing module with residual connections. CSMM uses different patches to obtain multi-scale features and uses deep separable convolution to learn the correlation between spatial dimensions and channels. Although the CSMM module can measure the importance of different channels and reduce the number of parameters, it ignores the information relationship between channels. To compensate for this loss, the outputs of different CSMM modules are combined through point-by-point 1x1 convolutions. The second part is a two-layer fully connected network that fuses the information of each channel and strengthens the connection between channels. The output logarithm learned by the fully connected layer will be processed by an exponential function to expand the value range from [0, 1] to [1, e]. This exponential normalization provides a monotonic mapping relationship, making the result more tolerant to position errors. Finally, the output of the SEAM module is multiplied with the original feature as attention to effectively deal with the problem of target occlusion.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A cable identification and counting method in a cable trench based on an improved YOLO model, characterized in that: The method comprises the following steps: Use the camera to collect image information in the cable trench; The collected image information is input into the improved YOLOv8 model for cable identification. The improved YOLOv8 model includes: The backbone network includes multiple low-level feature maps and high-level feature maps fusion (CSPDarknet53 to 2-Stage FPN, C2f) and characteristic receptive field fusion module (EFFVM), where the EFFVM module includes multiple convolution modules (Convoluton, CONv) with different expansion coefficients; The neck network part includes Feature Pyramid Networks (FPN) and Path Aggregation Network (PAN). The Concat connection in the FPN structure is replaced by the Fast Gradient Method (FGM) module. The FGM module includes multiple Enhanced guidance attention mechanisms (EGAMs). The head network part includes three detection heads, each of which includes a classification branch and a regression branch, and a new angle prediction output head is added; According to the cable identification results output by the improved YOLOv8 model, the number of cables is counted and output.

2. According to claim 1, a cable identification and counting method in a cable trench based on an improved YOLO model is characterized in that: The structure of the EFFVM module is as follows: Enter feature information; Feature information is processed through a CBS module; Connect three CONv modules with different expansion coefficients in series; Use Concat to concatenate the features processed by the three CONv modules with the input features; After a CONv module, feature fusion processing is performed; Output the fused feature information.

3. According to claim 1, a cable identification and counting method in a cable trench based on an improved YOLO model is characterized in that: The structure of the FGM module is as follows: Enter feature information; Use the same bidirectional EGAM module to provide feature guidance for each layer. Add and fuse the guidance feature information and the own feature information; Output the fused feature information.

4. According to claim 1, a cable identification and counting method in a cable trench based on an improved YOLO model is characterized in that: The structure of the dynamic snake convolution DSConv is as follows: Enter feature information; Use deformation offset △ to deform the convolution kernel; Perform convolution operation; Output the feature information after convolution.

5. According to claim 1, a cable identification and counting method in a cable trench based on an improved YOLO model is characterized in that: The structure of the SEAM module is as follows: Enter feature information; Feature extraction is performed through CSMM channels with residual connections and spatial mixing modules; The information of each channel is fused through a two-layer fully connected network; The output of the SEAM module is used as attention and multiplied with the original features; Output the feature information after attention enhancement.

6. According to claim 1, a method for identifying and counting cables in a cable trench based on an improved YOLO model is characterized in that: The improved YOLOv8 model uses the ProbIoU loss function to predict the rotated rectangular box; ProbIoU converts the real box and the predicted box into a Gaussian distribution ProbIoU designs a network that regresses three parameters (a′, b′, θ) to make the shape and rotation angle of the rotation prediction box closer to the real box; the similarity of the two Gaussian distributions is obtained by calculating the bus coefficient; The Hellinger distance is introduced to bring the two Gaussian distributions closer, so that the real box and the predicted box overlap more. The calculation formula of the Hellinger distance is as shown in formula (1), the calculation formula of ProbIoU is as shown in formula (2), and the final loss function formula is as shown in formula (3): Problem=1-H D (p,q) (2) L ProbIoU (p,q)=1-ProbIoU(p,q) (3) Among them, p(x) and q(x) represent the probability density functions of the Gaussian distribution of the real box and the predicted box respectively; B c (p,q) is the Bhattacharyya coefficient of the two probability density functions; H D (p,q) is the Hellinger distance between two probability density functions; ProbIoU is the Gaussian distribution similarity; L ProbIoU (p,q) is the positioning loss function.

7. According to claim 1, a method for identifying and counting cables in a cable trench based on an improved YOLO model is characterized in that: The improved YOLOv8 model uses dynamic snake convolution DSConv to extract local features of cable images.

8. According to claim 1, a method for identifying and counting cables in a cable trench based on an improved YOLO model is characterized in that: The improved YOLOv8 model uses a Spatially Enhanced Attention Module (SEAM) to enhance attention to the target area in the image and suppress the background area.

9. According to claim 1, a method for identifying and counting cables in a cable trench based on an improved YOLO model is characterized in that: The improved YOLOv8 model uses the EFFVM module to enhance the model's ability to recognize occluded targets.

Images