Electric transmission line nest detection method and device based on ER-YOLO

Abstract

The invention relates to an ER-YOLO-based power transmission line bird nest detection method and device, and in the bird nest detection task of a power transmission line, the shooting distance is difficult to unify, the bird nest has different sizes in a picture, the area of the picture occupied by the bird nest is extremely limited, and the detection background is complex and changeable. An original YOLOv10 network model often neglects feature information of a small target in a feature extraction process, so that the accuracy rate in bird nest target detection is low. In order to solve the problem, an efficient channel attention module is introduced into a YOLOv10 backbone network, and a Repulsion loss function is used. Through the improvements, the model can extract bird nest features more effectively, the influence of background noise on target detection is reduced by endowing effective channels with higher weights and inhibiting irrelevant background features, and finally, the average precision of a mean value on a data set reaches 98.9%.

Description

Technical Field

[0001] The present invention relates to the field of safe operation of distribution network equipment, and in particular to a method and device for detecting bird nests in power transmission lines based on ER (ECA-Repulsion)-YOLO. Background Art

[0002] The YOLO (You Only Look Once) series of algorithms have attracted much attention in the field of machine learning due to their real-time detection capabilities and high efficiency. As the latest research result of this series, YOLOv10 has made many innovations in model structure and training process, improving training efficiency and inference speed.

[0003] In the task of detecting bird nests on power transmission lines, it is difficult to unify the shooting distance, the bird nests have different sizes in the picture, and the area occupied by the bird nests is extremely limited. At the same time, the detection background is complex and changeable. The original YOLOv10 network model often ignores the feature information of small targets during feature extraction, which leads to low accuracy in bird nest target detection.

[0004] In summary, there are the following problems in the bird's nest detection task:

[0005] (1) Bird nests are usually small in size, and traditional deep learning algorithms often have poor performance in detecting small objects;

[0006] (2) Bird nests are often found on transmission line towers, which can easily lead to missed detections;

[0007] (3) The background of bird nest images is usually complex, which can easily lead to false detection.

[0008] The currently widely used object detection models are difficult to effectively solve these problems, especially for small-sized and occluded objects. Although some progress has been made in algorithm accuracy, the detection ability of small objects is still insufficient. Summary of the invention

[0009] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a transmission line bird nest detection method and device based on ER (ECA-Repulsion)-YOLO, so as to solve or partially solve the problem that the recognition accuracy of the YOLO model in the transmission line bird nest detection task is not ideal.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] In one aspect of the present invention, a method for detecting bird nests on power transmission lines based on ER-YOLO is provided. The image of the power transmission line to be detected is input into an ER-YOLO model pre-trained based on a repulsion loss function to obtain a bird nest detection result, wherein the ER-YOLO model includes a Backbone network, a Neck network and a Head network based on efficient channel attention, wherein the process of processing features by the Backbone network includes the following steps:

[0012] Step S1, for the transmission line image to be detected, firstly extract the basic features through convolution and C2f modules, and then obtain the intermediate features through stepwise down sampling and feature enhancement;

[0013] Step S2, using the intermediate features as input to an efficient channel attention module to extract channel global information;

[0014] Step S3, the output of the efficient channel attention module is processed by the SPPF and PSA modules in sequence to obtain a partial output of the Backbone network.

[0015] As a preferred technical solution, in step S2, the process of extracting channel global information using the efficient channel attention module includes:

[0016] For the intermediate features, channel weights are calculated through global average pooling and one-dimensional convolution, and the output of the efficient channel attention module is obtained by multiplying the intermediate features by the channel weights element by element.

[0017] As a preferred technical solution, the training process of the ER-YOLO model includes:

[0018] A data set is obtained and divided into a training set, a validation set, and a test set, and data enhancement processing is performed on the training set, and the ER-YOLO model is trained based on the enhanced training set.

[0019] As a preferred technical solution, the Repulsion loss function is calculated using the following formula:

[0020] L=L Attr +αL RepGT +βL RepBox

[0021] Among them, L is the repulsion loss, L Attr is an attraction term, which is used to limit the prediction box to be close to the bird's nest, L RepGT and L RepBox is the exclusion term, where L RepGT It is used to limit the predicted box to be away from the surrounding real box targets, LRepBox Used to limit the prediction box away from the prediction box of non-bird nest targets.

[0022] As a preferred technical solution, the bird's nest detection result includes the location of the bird's nest in the power transmission line image.

[0023] Another aspect of the present invention provides a power transmission line bird nest detection device based on ER-YOLO, comprising a Backbone network, a Neck network and a Head network that are sequentially connected and pre-trained based on a Repulsion loss function, and the power transmission line image to be detected passes through each network in turn to obtain a bird nest detection result, wherein the Backbone network comprises a basic feature extraction module, an intermediate layer feature extraction module, an efficient channel attention module, an enhanced model's expression capability and a module for focusing on key features that are sequentially connected.

[0024] As a preferred technical solution, the efficient channel attention module includes:

[0025] A global average pooling layer, used to process the output of the intermediate layer feature extraction module;

[0026] A one-dimensional convolution layer, used to perform a convolution operation on the output of the global average pooling layer to obtain channel weights;

[0027] The output layer is used to perform principal element multiplication on the output of the intermediate layer feature extraction module and the channel weight to obtain the output of the efficient channel attention module.

[0028] As a preferred technical solution, the basic feature extraction module includes a first convolutional layer, a second convolutional layer, a first C2f module and a third convolutional layer connected in sequence.

[0029] As a preferred technical solution, the intermediate layer feature extraction module includes a second C2f block, a first SCDown block, a third C2f block, a second SCDown block and a fourth C2f block connected in sequence.

[0030] As a preferred technical solution, the module for enhancing the expressiveness of the model and focusing on key features includes a SPPF block and a PSA block connected in sequence.

[0031] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0032] (1) Strong ability to recognize bird nests: The present invention combines YOLO and the efficient channel attention module (ECA-NET) to weight the original feature map, so that important channel features can be highlighted, while effectively suppressing the interference of irrelevant background information. This mechanism of enhancing feature extraction capability enables ER to perform better in tasks such as bird nest detection that require processing small targets and complex backgrounds.

[0033] (2) Reducing the interference of noise on bird nest recognition: The present invention combines the Repulsion loss function to train the ER-YOLO model. By introducing a repulsion mechanism in model training, the detection frame is encouraged to stay away from background objects that are not related to the target, thereby reducing the impact of background interference on the detection results. The detection frame is more focused on the bird nest itself, which improves the positioning accuracy of the model in complex scenes and effectively improves the overall performance of the bird nest detection task. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a schematic diagram of the ER-YOLO model in the embodiment;

[0035] Figure 2 Schematic diagram of the ER channel attention module in the embodiment;

[0036] Figure 3 Schematic diagram of data enhancement results in the embodiment. DETAILED DESCRIPTION

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0038] Example 1

[0039] In view of the problems existing in the aforementioned prior art, this embodiment provides a transmission line bird nest detection method based on ER (ECA-Repulsion)-YOLO. This method makes targeted improvements based on the YOLOv10 target detection model, adopts the ECA (Efficient Channel Attention) module to improve the small target recognition ability, combines the Repulsion loss function to reduce the interference of the background on the detection results, and improves the generalization ability of the model through data enhancement, which can more effectively solve the bird nest detection problem.

[0040] YOLOv10 has made many innovations in model structure and training process. First, by introducing a consistent dual allocation strategy without non-maximum suppression (NMS), the training efficiency and inference speed are significantly improved. Secondly, YOLOv10 adopts a comprehensive efficiency and accuracy driven design strategy, which not only optimizes each component of the network, but also greatly reduces the computational overhead and improves model performance. This embodiment uses YOLOv10 as the detection framework in order to achieve accurate recognition of bird nests while ensuring real-time performance.

[0041] This method inputs the transmission line image to be detected into the ER-YOLO model pre-trained based on the Repulsion loss function to obtain the bird's nest detection result. Specifically, the bird's nest detection result includes the location of the bird's nest in the transmission line image. The ER-YOLO model includes a Backbone network (i.e., backbone network), a Neck network, and a Head network based on efficient channel attention. The process of Backbone network processing features includes the following steps:

[0042] Step S0, training the ER-YOLO model, obtaining a data set and dividing it into a training set, a validation set, and a test set, and performing data augmentation processing on the training set, with the goal of minimizing the loss function including the Repulsion loss, and training the ER-YOLO model based on the enhanced training set.

[0043] Specifically, data enhancement methods include rotation, translation, scaling, flipping, brightness adjustment, contrast adjustment, and noise addition.

[0044] In order to solve the problem of unsatisfactory accuracy in detecting bird nests in complex backgrounds, this embodiment uses the Repulsion loss function to train the model. Considering that bird nests are usually located in complex environments such as power lines and tree branches, these background objects can easily interfere with the detection model, resulting in false detection or missed detection. In order to solve this problem, the loss function introduces a "repulsion" mechanism in model training to encourage the detection frame to stay away from background objects that are not related to the target, thereby reducing the impact of background interference on the detection results. This method makes the detection frame more focused on the bird nest itself, improves the positioning accuracy of the model in complex scenes, and effectively improves the overall performance of the bird nest detection task. The loss function is:

[0045] L=L Attr +αL RepGT +βL RepBox

[0046] Among them, L is the repulsion loss, L Attr is an attraction term, which is used to limit the prediction box to be close to the bird's nest, L RepGT and L RepBox is the exclusion term, where L RepGTIt is used to limit the predicted box to be away from the surrounding real box targets, L RepBox Used to limit the prediction box away from the prediction box of non-bird nest targets.

[0047] Step S1, for the transmission line image to be detected, firstly extract the basic features through convolution and C2f modules, and then obtain the intermediate features through step-by-step down sampling and feature enhancement.

[0048] In step S2, the intermediate features are used as input to the efficient channel attention module to extract the channel global information.

[0049] Specifically, for the intermediate features, the channel weights are calculated through global average pooling and one-dimensional convolution, and the output of the efficient channel attention module is obtained by multiplying the intermediate features by the channel weights element by element.

[0050] The Efficient Channel Attention (ECA) module is designed to enhance the channel feature extraction capability of deep learning models. Figure 2 As shown in Figure 2. The core process of the ECA module first extracts the feature map F[h,w,c] from the backbone network, where h, w, and c represent the height, width, and number of channels of the feature map, respectively. Next, the average value of each channel is calculated through the global average pooling (GAP) operation to generate a new feature map of size [1×1×C]. This process effectively extracts the global information of the channel.

[0051] This module uses a one-dimensional convolution operation of size 5 instead of a fully connected layer when generating channel weights. This design avoids the problem of dimensionality reduction and helps retain more feature information, while reducing the number of parameters and improving the computational efficiency of the model. ECA weights the original feature map to highlight important channel features while effectively suppressing the interference of irrelevant background information. This mechanism of enhancing feature extraction capabilities enables ECA to perform better in tasks such as bird nest detection that require processing small targets and complex backgrounds.

[0052] Step S3: The output of the efficient channel attention module is processed by the SPPF and PSA modules in turn to obtain a partial output of the Backbone network.

[0053] In order to verify the effectiveness of the bird nest detection algorithm ER-YOLO proposed in this paper, this paper compares the algorithm with the YOLOv10 algorithm. The hardware and software conditions used are shown in Table 1.

[0054] Table 1 Experimental software and hardware parameters

[0055]

[0056]

[0057] Specifically, the dataset used in the comparative experiment contains 1,000 images. In order to improve the generalization ability of the model, the dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. On this basis, in order to further enhance the diversity and robustness of the dataset, we performed data augmentation on the training set. The images before and after data augmentation are shown in Figure 2. Figure 3 shown.

[0058] In the bird nest detection task, this study uses precision (Precision, P), recall (Recall, R), average precision (Average Precision, AP) and mean average precision (Mean Average Precision, mAP) as the main indicators for evaluating the detection accuracy of the model. Specifically, AP@0.5 is used to measure the detection accuracy of the bird nest target when the intersection over Union (IoU) is 0.5; while mAP@0.5 represents the average detection accuracy of all bird nest samples under the same IoU threshold. Furthermore, mAP@0.5:0.95 considers the average accuracy of bird nest detection under different IoU thresholds (from 0.5 to 0.95, with a step size of 0.05).

[0059] The relevant calculation formula is:

[0060]

[0061] In the formula, TP means the model predicts the correct result, FP means the model incorrectly predicts a positive result, and FN means the model mistakenly predicts a negative result as a positive result.

[0062] The experimental results are shown in Table 2.

[0063]

[0064] The experimental results show that although the reasoning time of the ER-YOLO algorithm is slightly increased, it is improved in all indicators compared with YOLOv10. The application scenario of bird nest detection on power transmission lines usually requires a higher accuracy, and for the reasoning time, it only needs to be able to ensure real-time execution. Therefore, the ER-YOLO algorithm provided in this embodiment is better than the YOLOv10 algorithm in bird nest detection.

[0065] In order to verify the actual use effect of the ER-YOLO algorithm, a test experiment was conducted using a pre-defined test set of 100 pictures. The experimental results show that the ER-YOLO algorithm detects all bird nest targets in the test pictures and can better meet the use requirements of actual scenarios.

[0066] In summary, in the bird's nest detection task of the transmission line, since the shooting distance is difficult to unify, the bird's nest has different sizes in the picture, and the area occupied by the bird's nest is extremely limited, and the detection background is complex and changeable. The original YOLOv10 network model often ignores the feature information of small targets in the feature extraction process, which leads to low accuracy in bird's nest target detection. In order to solve this problem, this embodiment introduces an efficient channel attention module into the YOLOv10 backbone network and uses the Repulsion loss function. These improvements enable the model to extract bird's nest features more effectively, by giving higher weights to effective channels and suppressing irrelevant background features, thereby reducing the impact of background noise on target detection, and finally the mean average accuracy on the data set reached 98.9%.

[0067] Example 2

[0068] Based on Example 1, this example provides a transmission line bird nest detection device based on ER-YOLO, see Figure 1 , including the Backbone network, Neck network and Head network that are connected sequentially and pre-trained based on the Repulsion loss function. The image of the power transmission line to be detected passes through each network in turn to obtain the bird's nest detection result. Among them, the Backbone network includes the sequentially connected basic feature extraction module, the intermediate layer feature extraction module, the efficient channel attention module, the expression ability of the enhanced model and the attention module to key features.

[0069] Among them, the efficient channel attention module includes:

[0070] The global average pooling layer is used to process the output of the feature extraction module in the middle layer;

[0071] One-dimensional convolution layer, used to perform convolution operation on the output of the global average pooling layer to obtain channel weights;

[0072] The output layer is used to perform principal element multiplication on the output of the intermediate layer feature extraction module and the channel weight to obtain the output of the efficient channel attention module.

[0073] The feature extraction module includes a first convolutional layer, a second convolutional layer, a first C2f module and a third convolutional layer which are connected in sequence.

[0074] Among them, the middle-layer feature extraction module includes a second C2f block, a first SCDown block, a third C2f block, a second SCDown block and a fourth C2f block connected in sequence.

[0075] Among them, the module for enhancing the expressiveness of the model and focusing on key features includes sequentially connected SPPF blocks and PSA blocks.

[0076] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for detecting bird nests on power transmission lines based on ER-YOLO, characterized in that: The transmission line image to be detected is input into the ER-YOLO model pre-trained based on the Repulsion loss function to obtain the bird's nest detection result, wherein the ER-YOLO model includes a Backbone network, a Neck network and a Head network based on efficient channel attention, wherein the Backbone e The process of network feature processing includes the following steps: Step S1, for the transmission line image to be detected, firstly extract the basic features through convolution and C2f modules, and then obtain the intermediate features through stepwise down sampling and feature enhancement; Step S2, using the intermediate features as input of an efficient channel attention module to extract channel global information; Step S3, for the output of the efficient channel attention module, pass through the SPPF and PSA modules in turn to obtain the Backbon e Partial output of the network.

2. According to the ER-YOLO-based transmission line bird nest detection method of claim 1, it is characterized in that: In step S2, the process of extracting channel global information using the efficient channel attention module includes: For the intermediate features, channel weights are calculated through global average pooling and one-dimensional convolution, and the output of the efficient channel attention module is obtained by multiplying the intermediate features by the channel weights element by element.

3. The method for detecting bird nests on power transmission lines based on ER-YOLO according to claim 1, characterized in that: The training process of the ER-YOLO model includes: A data set is obtained and divided into a training set, a validation set, and a test set, and data enhancement processing is performed on the training set, and the ER-YOLO model is trained based on the enhanced training set.

4. The method for detecting bird nests on power transmission lines based on ER-YOLO according to claim 1, characterized in that: The repulsion loss function is calculated using the following formula: L=L Attr +αL RepGT +βL RepBox Among them, L is the repulsion loss, L Attr is an attraction term, which is used to limit the prediction box to be close to the bird's nest, L RepGT and L RepBox is the exclusion term, where L RepGT It is used to limit the predicted box to be away from the surrounding real box targets, L RepBox Used to limit the prediction box away from the prediction box of non-bird nest targets.

5. The method for detecting bird nests on power transmission lines based on ER-YOLO according to claim 1, characterized in that: The bird's nest detection result includes the location of the bird's nest in the power transmission line image.

6. A transmission line bird nest detection device based on ER-YOLO, characterized in that: The method comprises a Backbone network, a Neck network and a Head network which are sequentially connected and pre-trained based on the Repulsion loss function. The image of the power transmission line to be detected is passed through each network in turn to obtain the bird's nest detection result. The Backbone network comprises a sequentially connected basic feature extraction module, an intermediate layer feature extraction module, an efficient channel attention module, an expression ability of the enhanced model and a module for focusing on key features.

7. The ER-YOLO-based transmission line bird nest detection device according to claim 6, characterized in that: The efficient channel attention module includes: A global average pooling layer, used to process the output of the intermediate layer feature extraction module; A one-dimensional convolution layer, used to perform a convolution operation on the output of the global average pooling layer to obtain channel weights; The output layer is used to perform principal element multiplication on the output of the intermediate layer feature extraction module and the channel weight to obtain the output of the efficient channel attention module.

8. The ER-YOLO-based transmission line bird nest detection device according to claim 6, characterized in that: The basic feature extraction module includes a first convolutional layer, a second convolutional layer, a first C2f module and a third convolutional layer which are connected in sequence.

9. The ER-YOLO-based transmission line bird nest detection device according to claim 6, characterized in that: The intermediate layer feature extraction module includes a second C2f block, a first SCDown block, a third C2f block, a second SCDown block and a fourth C2f block connected in sequence.

10. The ER-YOLO-based transmission line bird nest detection device according to claim 6, characterized in that: The module for enhancing the expressiveness of the model and focusing on key features includes a SPPF block and a PSA block connected in sequence.

Images