A power transmission line bird identification method based on deep learning

Abstract

The application provides a kind of power transmission line bird identification method based on deep learning, it is related to the technical field of bird damage detection of power transmission line.The application converts monitoring video frame into pixel color value array and inputs target detection network, extracts multi-scale feature map using backbone feature, and realizes attention enhancement and multi-scale feature fusion through improved feature fusion network containing scale sequence attention module SSA and down-sampling feature fusion module ADFF, generates fusion feature map corresponding to P2 to P5 scale.Again, by perception detection head module, boundary box regression and class prediction are executed, and bird class and position results are output.Training phase is based on bird image with real label to construct training sample set, and PSIoU-based PSLoss loss function is used to iteratively update network parameters.The application can realize high-precision and strong-robustness identification of multi-scale bird targets in complex line scene.

Description

Technical Field

[0001] This invention relates to the field of bird damage detection technology for power transmission lines, and in particular to a deep learning-based method for bird identification on power transmission lines. Background Technology

[0002] Transmission lines are exposed to the outdoor environment for extended periods, attracting frequent bird activity, which significantly impacts the safe operation of power grid equipment. Statistics show that bird-related line tripping incidents rank third among all power grid equipment tripping accidents in the State Grid, second only to lightning strikes and external damage. Some birds nest on poles, and the nesting materials can cause short circuits or insulation failures. Large birds flying near overhead lines can also cause momentary ground short circuits or phase-to-phase discharge hazards, seriously affecting the stability and reliability of the power system. Therefore, accurate and real-time detection of birds on transmission lines is of significant engineering importance.

[0003] Currently, bird damage monitoring in power grids mainly relies on methods such as voiceprint recognition, Doppler radar, dual-spectral imaging, and manual inspection. However, each method has limitations, including noise sensitivity, inability to distinguish species, high equipment costs, and low inspection efficiency. With the development of deep learning target detection technology, YOLO series models are gradually being applied to bird identification scenarios. For example... Figure 1 As shown, YOLOv8, as a relatively new detection framework, consists of a Backbone, Neck, and Detect module. The Backbone extracts multi-scale features of different receptive fields through Conv, C2f, and spatial pyramid pooling. The Neck adopts a top-down and bottom-up feature fusion structure to integrate high-level semantic features with low-level detail features, generating a fused feature map of three scales to meet the needs of large, medium, and small target detection. The multi-scale features output by the Neck are finally input into the Detect module to complete target prediction.

[0004] While YOLOv8 possesses strong general object detection capabilities, it still has significant shortcomings in power transmission line scenarios: complex backgrounds result in images containing a large amount of redundant information such as towers, conductors, and ground environment, complicating the feature extraction process; bird targets are small, dense, and occluded, leading to a significant decrease in detection accuracy; and the large scale variation and complex imaging angles of targets in power transmission line scenarios make multi-scale detection tasks more difficult. Based on these issues, existing deep learning detection methods still struggle to meet the engineering requirements of high accuracy and robustness in bird recognition scenarios on power transmission lines. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a deep learning-based method for bird identification on power transmission lines. By introducing a scale sequence attention module (SSA), a downsampled feature fusion module (ADFF), and a PSIoU-based PSLoss loss function, this method achieves high-precision and robust identification of multi-scale bird targets in complex power transmission line scenarios.

[0006] To achieve the above objectives, the present invention provides the following solution: A deep learning-based method for bird identification on power transmission lines includes: The video from the power transmission line monitoring camera is collected, and the video is split into video frame images by frame, and each video frame image is converted into a pixel color value array; The pixel color value array is input into the backbone feature extraction part of the target detection network to extract a multi-scale feature map. The multi-scale feature maps are input into the improved feature fusion network in the target detection network to perform attention enhancement and multi-scale feature fusion on the multi-scale feature maps, resulting in fused feature maps corresponding to the four scales P2 to P5. The improved feature fusion network includes a Scale Sequence Attention (SSA) module and a Downsampling Feature Fusion (ADFF) module. The SSA module receives the multi-scale feature maps, enhances the weights of bird target-related features, and suppresses redundant background features. The ADFF module receives the features processed by the SSA module at each scale and performs feature fusion between adjacent scales to generate fused feature maps corresponding one-to-one with the detection layers P2 to P5. The fused feature map is input into the perception detection head module, which performs perception enhancement on the fused feature map at each scale and performs bounding box regression and category prediction to obtain candidate bounding boxes and the confidence scores of bird categories corresponding to each candidate bounding box. During the training phase, a training sample set is constructed based on bird images with real annotations. The training sample set is then used as a supervision signal input to the target detection network, and the network parameters are iteratively updated using the PSLoss loss function. After the PSLoss loss function converges, a bird detection model is obtained. The video frame images collected during the process are converted into the pixel color value array and input into the bird detection model. The output is the transmission line bird detection result with category label and location coordinates.

[0007] Preferably, converting each of the video frame images into an array of pixel color values ​​includes: Each video frame image is divided into multiple pixels; Each pixel's continuous color value is quantized into an integer between 0 and 255, and the pixel's position in the video frame image and its corresponding color channel value are organized as array elements to form the pixel color value array.

[0008] Preferably, the backbone feature extraction part of the target detection network includes a convolutional structure, a C2f structure, and a spatial pyramid pooling module connected in sequence. When the pixel color value array is input into the backbone feature extraction part of the target detection network, multi-level image features are first extracted through the convolutional structure and the C2f structure, and then the features of different receptive fields are aggregated through the spatial pyramid pooling module to improve the adaptability to bird targets of different sizes.

[0009] Preferably, the Scale Sequence Attention (SSA) module performs a global pooling operation on the input multi-scale feature map, compressing the spatial information of each channel into a global feature value. Then, it sequentially passes through a first fully connected layer and a nonlinear transformation with a ReLU activation function to perform dimensionality transformation and preliminary feature selection. Next, it passes through a second fully connected layer and a nonlinear transformation with a Sigmoid activation function to map the feature value to the range of 0 to 1 to generate channel-level attention weights. The channel-level attention weights are multiplied element-wise with the original features and added to the original features through an addition operation, thereby strengthening the important channel features related to bird target recognition and suppressing redundant background features.

[0010] Preferably, the downsampling feature fusion module ADFF takes large-scale features, medium-scale features, and small-scale features as inputs. It performs max pooling and convolution on the large-scale features to achieve downsampling, average pooling and convolution on the medium-scale features to extract medium-scale features, and directly convolution on the small-scale features to extract fine-grained features, thereby obtaining three feature maps with the same spatial size. The three feature maps with the same spatial size are then connected in the channel dimension to generate a fusion feature map for identifying densely overlapping bird targets.

[0011] Preferably, the improved feature fusion network further includes an additional P2 detection layer specifically designed for small target features, disposed in the neck structure of the target detection network. The additional P2 detection layer is connected to an improved scale sequence feature fusion structure composed of two downsampled feature fusion modules ADFF and two scale sequence attention modules SSA, so that the fused feature maps corresponding to the four scales P2 to P5 correspond one-to-one with the P2 to P5 detection layers, thereby improving the detection capability of dense small target birds.

[0012] Preferably, in the fusion feature maps corresponding to the four scales P2 to P5, the fusion feature map at scale P2 is generated by the additional P2 detection layer. The resolution of the fusion feature map at scale P2 is 160×160 pixels. It is used to retain the semantic information related to small-scale bird targets under the 160×160 pixel resolution condition and to achieve accurate detection and localization of the small-scale bird targets. It works in conjunction with the detection layers corresponding to scales P3, P4 and P5 to improve the detection sensitivity of the small-scale bird targets while maintaining the ability to recognize large-scale bird targets.

[0013] Preferably, the perception detection head module includes a perception module, which is configured to receive the fused feature map at four scales from P2 to P5, and sequentially perform a 1×1 kernel-based group normalization convolution operation and the processing of the perception module on the fused feature map at each scale; the perception module includes: The first perceptual branch, which consists of average pooling, 1×1 convolution, ReLU activation unit and Sigmoid activation unit in sequence, is used to generate scale attention weights and weight the fused feature maps at different scales. The second perception branch, which consists of deformable convolution, 3×3 convolution and sigmoid activation unit in sequence, is used to generate spatial attention weights and enhance the local spatial features of occluded bird targets and small-scale bird targets. The third perception branch consists of average pooling, a fully connected layer, a ReLU activation unit, a fully connected layer, and a normalization operation Normalize in sequence. The normalization operation Normalize is implemented by a hard sigmoid function to limit the output value to the range of 0 to 1, thereby normalizing the perception features processed by the perception module.

[0014] Preferably, the PSLoss loss function is a total loss function constructed based on the PSIoU index. The PSLoss loss function includes a distribution focus loss term for category prediction, a PSIoU-based bounding box loss term for bounding box regression, and a binary cross-entropy loss term for target presence determination. Specifically, the PSIoU-based bounding box loss term introduces a scaling factor (scale) and a shape cost (Ω) when calculating the overlap between the predicted and true bounding boxes. The scaling factor (scale) is related to the size of the targets in the dataset and is used to control the scale of the auxiliary bounding boxes. The shape cost (Ω) characterizes the shape differences of the bounding boxes and, combined with the width and height of the true bounding boxes, the distance between the center points of the predicted and true bounding boxes, and the positional relationships of the left, right, top, and bottom boundaries of the predicted and true bounding boxes, to improve the accuracy of position regression for small-scale bird targets and accelerate the convergence of the loss function.

[0015] Preferably, the training sample set includes bird images from the open-source bird image dataset CUB-200 and additionally collected images of common birds. The open-source bird image dataset CUB-200 includes 11,788 images of 200 different bird species. During training, the input data of the training sample set is preprocessed to a uniform resolution of 640×640 pixels. Without loading pre-trained weights, the iteration period is set to 200 times, the batch size is 16, the initial base learning rate is 0.01, and the weight decay coefficient is 0.0005 to improve the robustness of the bird detection model in complex power transmission line scenarios.

[0016] The present invention discloses the following technical effects: This invention overcomes the shortcomings of voiceprint recognition, such as its reliance on calls, sensitivity to noise, and difficulty in capturing stationary birds, by converting continuous video frames captured by power line monitoring cameras into pixel color value arrays and inputting them into a deep learning model for end-to-end inference. It achieves stable recognition capabilities independent of sound and unaffected by environmental noise. Compared to Doppler radar's inability to distinguish between birds and insects and its difficulty in achieving species-level identification, this invention, based on convolutional features and a detection head's category prediction mechanism, can output a clear bird target category, improving species-level identification accuracy.

[0017] This invention utilizes existing video surveillance resources on power transmission lines for identification, eliminating the need for additional high-cost infrared dual-spectrum equipment, manual tower climbing inspections, or reliance on personnel experience, significantly reducing system construction and maintenance costs. Through automated reasoning using deep learning models, this invention enables all-weather, real-time bird detection. Compared to manual inspections, which are characterized by long cycles, high missed detection rates, and risks associated with working at heights, this invention offers advantages in terms of continuity and reliability.

[0018] This invention introduces a scale sequence attention module (SSA) in the feature fusion stage. By using channel attention weights, it suppresses a large number of tower and wire background features in transmission line images, while strengthening salient features related to birds. This effectively solves the problem of "weak target saliency and increased false detection rate" in traditional YOLOv8 in complex line backgrounds, thereby improving the model's feature recognition ability in complex scenes.

[0019] This invention employs the Downsampling Feature Fusion (ADFF) module to achieve cross-layer fusion of multi-scale features and adds a dedicated P2 detection layer for small target detection. This allows for a more thorough fusion of deep semantic features and shallow spatial details, significantly enhancing the model's ability to detect small-scale birds, distant birds, and densely occluded birds. Compared to the low recall rate of the original YOLOv8 for small targets, this invention achieves a higher detection rate and more stable localization accuracy.

[0020] This invention utilizes the PSLoss loss function based on PSIoU, incorporating scaling factors and shape cost terms into the bounding box regression process. It simultaneously considers category prediction and foreground determination, achieving stable regression optimization for bird targets at different scales. This loss function converges faster and exhibits more stable gradients during training, significantly improving the localization accuracy of small targets and overcoming the limitation of traditional IoU-based losses in handling large scale differences. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments 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.

[0022] Figure 1 This is a diagram of the network structure in the existing technology; Figure 2 A flowchart of the method provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the SSA structure provided in an embodiment of the present invention; Figure 4 A schematic diagram of the ADFF module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the Phead detection head provided in an embodiment of the present invention; Figure 6 A schematic diagram of the sensing module provided in an embodiment of the present invention; Figure 7 A schematic diagram of the improved YOLO structure provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of actual test results provided for an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The purpose of this invention is to provide a deep learning-based method for bird identification on power transmission lines, which can significantly improve the detection of small target birds, identification of obstructed environments, and suppression of false detections without increasing hardware costs, providing an efficient, deployable, and scalable intelligent monitoring method for preventing bird damage to power grids.

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0026] Figure 2 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a deep learning-based method for bird identification on power transmission lines, comprising: Step 100: Acquire video from the power transmission line monitoring camera, split the video into video frame images, and convert each video frame image into a pixel color value array; Step 200: Input the pixel color value array into the backbone feature extraction part of the object detection network to extract multi-scale feature maps; Step 300: Input the multi-scale feature map into the improved feature fusion network in the target detection network to perform attention enhancement and multi-scale feature fusion on the multi-scale feature map to obtain fused feature maps corresponding to the four scales P2 to P5; wherein, the improved feature fusion network includes a scale sequence attention module SSA and a downsampling feature fusion module ADFF. The scale sequence attention module SSA receives the multi-scale feature map, performs weight enhancement on bird target-related features and suppresses redundant background features. The downsampling feature fusion module ADFF receives the features of each scale after being processed by the scale sequence attention module SSA, performs feature fusion between adjacent scales, and generates fused feature maps that correspond one-to-one with the detection layers P2 to P5; Step 400: Input the fused feature map into the perception detection head module. The perception detection head module performs perception enhancement on the fused feature map at each scale and performs bounding box regression and category prediction to obtain candidate bounding boxes and the confidence scores of bird categories corresponding to each candidate bounding box. Step 500: During the training phase, a training sample set is constructed based on bird images with real annotations. The training sample set is used as a supervision signal input to the target detection network, and the network parameters are iteratively updated using the PSLoss loss function. Step 600: After the PSLoss loss function converges, the bird detection model is obtained. The video frame images collected during the process are converted into pixel color value arrays and input into the bird detection model. The output is the bird detection results of the transmission line with category labels and location coordinates.

[0027] The network structure in this embodiment is an improvement upon the current YOLOv8 algorithm. In the current YOLOv8 algorithm, the Backbone is responsible for extracting multi-scale features from the input image, providing basic feature representations for subsequent detection. The input image sequentially passes through Conv and C2f, and finally through a spatial pyramid pooling module to aggregate features from different receptive fields, improving adaptability to targets of different sizes. Neck integrates features from different levels through "top-down + bottom-up" feature fusion, enhancing the detection capability of multi-scale targets and generating three fused feature maps at different scales, corresponding to the detection needs of large, medium, and small targets respectively. This allows for full fusion of high-level and low-level features, compensating for the shortcomings of single-scale features. The three scale feature maps output by Neck are input into the Detect module, each performing the detection of targets at its corresponding scale.

[0028] In this embodiment, the video captured by the camera is converted into frame-by-frame images. These frames are then divided into H×W pixels, and the continuous color values ​​of each pixel are converted into integers from 0 to 255. The "pixel position" and "color channel value" are mapped to elements of a multidimensional array. This array is then input into the improved YOLOv8 network. The modified modules are as follows: 1. Improved feature fusion module Feature fusion networks can integrate feature information from various levels, enabling effective detection of multi-scale targets. Successive downsampling and feature extraction are used to increase the semantic richness of features while reducing noise levels. However, this leads to a reduction in the scale of the output feature map, so feature information for small objects may be partially lost after successive downsampling. Top-down feature pyramid networks are almost incapable of integrating features of small objects. Furthermore, deep feature maps acquire a greater amount of semantic information, and spatial details are lost as network depth increases, affecting target detection accuracy. To address this issue, this embodiment redesigns the feature fusion network structure, which consists of five parts: an additional P2 detection layer specifically for small target features, two ADFF modules, and two SSA modules for improved scale sequence feature fusion.

[0029] Figure 3The diagram shows the Scaled Sequence Attention (SSA) module. Unsqueeze and Squeeze are used to increase and decrease data dimensionality, respectively. `concat` is used to join multiple data tensors along a specified dimension. Global pooling performs a global pooling operation on the input feature map, compressing the spatial information of each channel into a single global feature value, thereby capturing global contextual information. The fully connected layer (FC) inputs the globally pooled features into the first fully connected layer, performing dimensionality transformation and initial nonlinear mapping. The ReLU activation function introduces nonlinearity, enhancing the model's expressive power while filtering out some irrelevant information. The features then pass through another fully connected layer to further adjust the feature dimensionality, preparing for subsequent attention weight generation. The Sigmoid activation function maps feature values ​​to between 0 and 1, generating channel-level attention weights that reflect the importance of each channel. The Add operation multiplies the attention weights element-wise with the original features and then adds them back to the original features, strengthening the features of important channels and ultimately outputting attention-enhanced features.

[0030] The attention module generates channel-level attention weights through global information compression and nonlinear transformation, dynamically enhancing key channel features for target recognition while suppressing background or redundant information. Features at each scale are first extracted through convolution to obtain basic features, and then the attention module enhances key information to prepare for subsequent stitching.

[0031] After concatenating multi-scale features, the results are integrated via convolution, reduced by max pooling, and finally fused using Squeeze (dimensionality compression). This module simultaneously processes small, medium, and large-scale features, capturing both the details and local structure of the target while also grasping its global outline. It is adaptable to targets of different sizes and shapes, such as multi-scale recognition of birds, from hummingbirds to eagles. To identify densely overlapping bird species, a downsampled feature fusion module (ADFF) was designed, such as... Figure 4 As shown, ADFF's input consists of large-scale, medium-scale, and small-scale features. The input data is first subjected to average pooling to initially reduce data dimensionality, decrease computational cost, and preserve overall feature trends. For large-scale feature maps, max pooling and convolution are combined for downsampling and preserving high-resolution features and details. Medium-scale features are first average pooled to further reduce dimensionality before convolution to extract medium-scale features. For small-scale features, convolution is directly applied to the chunk output features to extract fine-grained small-scale features. Finally, the three feature maps of the same size are concatenated along the channel dimension.

[0032] The initial YOLOv8 network used three detector heads to process feature maps at different scales: 80×80, 40×40, and 20×20. Larger resolution feature maps have smaller receptive fields and provide detailed local features and location information, which helps in recognizing small objects. In contrast, smaller feature maps have larger receptive fields and capture richer semantic information, increasing their applicability to larger object recognition. However, due to multiple downsampling, small object features become less sharp, sometimes even blurry against the background. Therefore, an additional feature layer P2, enriched with semantic information about small objects, was added to the neck of the model. The 160×160 resolution facilitates accurate detection and localization of bird targets within deeper feature maps.

[0033] By integrating surface feature information into a deep semantic layer that encodes global information, the generated model's ability to retain features relevant to small targets is improved. The collaborative work of four recognition heads enhances the model's resolution and sensitivity to small target detection, helping it achieve more comprehensive recognition performance.

[0034] 2. Improvements to the detection head Since each detection head has independent feature inputs, there is a lack of information exchange between heads, which reduces detection performance. Furthermore, the introduction of the P2 small target detection layer increases the number of floating-point operations in the model to some extent. To address these issues, this embodiment designs a perceptual detection head (Phead), and the improved detection head structure is as follows: Figure 5 As shown.

[0035] After receiving feature input from layers P2-P5, Phead first applies group-normalized convolutions with a kernel size of 1×1 to each level. This step facilitates cross-channel information exchange, enriches target information, and improves the classification and regression performance of the detection head. Then it enters the perception module, such as... Figure 6 As shown.

[0036] The first part consists of average pooling, 1x1 convolution, ReLU+h, and sigmoid. Scale attention weights are generated through global pooling and nonlinear transformations to dynamically adjust the contribution of input features at different scales.

[0037] The middle part consists of deformable convolutions, 3x3 convolutions, and sigmoid functions. Spatial convolutions generate spatial attention and deformation information, dynamically focusing on the spatial region of the target. For example, it enhances the local spatial features of occluded or small targets, improving the accuracy of position perception.

[0038] The final part consists of average pooling + fully connected layer + ReLU + fully connected layer + Normalize. Normalize here is implemented using hardsigmoid, with the specific formula as follows: x represents the input data for the Normalize layer. To make... The output value is greater than 0 and less than 1, so use the maximum value function and the minimum value function.

[0039] 3. PSLoss Loss Function YOLOv8's original loss function L as follows: in It is distributed focal loss. It is the bounding box loss. It is a binary cross-entropy loss, and these three losses are defined as follows: in, y represents the probability score corresponding to the adjacent discrete anchor points predicted by the model; y represents the true continuous target value of the regression task. y represents the values ​​of two adjacent discrete anchor points containing the true value y; q represents the intersection-union ratio (IU / I) between the predicted bounding box and the true bounding box. This represents the square of the Euclidean distance between the center point of the predicted bounding box and the center point of the ground truth bounding box; represents the square of the diagonal length of the smallest bounding box that can simultaneously contain both the predicted and ground truth boxes; v represents a parameter that measures the consistency of the aspect ratio between the predicted and ground truth boxes. This represents the true label of the i-th sample; This represents the probability that the model predicts the i-th sample belongs to the positive class; n is the total number of samples.

[0040] Because birds move frequently and their target scale changes drastically, models struggle to pinpoint their location accurately. The original YOLOv8 network uses the CIoU loss function, which primarily focuses on shape loss and is less sensitive to positional deviations of small targets. In contrast, we propose a more accurate loss function calculation method, PSIoU, which more effectively considers the inherent properties of the bounding box, such as shape and scale, and their impact on regression. This allows for more effective handling of targets of different shapes and sizes, improving its performance in detecting small targets. The overall loss function... L As shown below: in The derivation formula is as follows: In the formula, scale This represents the scaling factor, which is related to the size of the target in the dataset. The width and height of the ground truth bounding box are represented as follows: w gt and h gt . v and h These represent the weighting coefficients in the horizontal and vertical directions, respectively. `distance` represents the distance loss. `Ω` represents the shape cost. The value of `θ` defines the shape cost, typically set to 4. This embodiment introduces a scaling factor to control the size of the auxiliary bounding box. This factor further accelerates convergence, balances the localization accuracy between targets of different scales, and improves generalization performance. b l , b r , b t , b b These represent the left, right, top, and bottom boundaries of the prediction box, respectively. Additionally, let... , , , These represent the left, right, top, and bottom boundaries of the ground truth bounding box, respectively. The scaling factor is represented by `ratio`, with values ​​ranging from 0.5 to 1.5. The center point of the ground truth bounding box is represented by (...). , ), ( x c , y c ) represents the center point of the pre-anchor box. i represents the area of ​​the intersection between the predicted box and the ground truth box, and u represents the area of ​​the non-overlapping portion of the predicted box and the ground truth box.

[0041] As an optional implementation, the training dataset in this embodiment comes from the open-source dataset CUB-200, which includes 200 different bird species and a total of 11,788 images. Common bird images are also included to increase robustness and help the model learn to distinguish the feature differences between the background and the actual target, thereby reducing misjudgments of complex backgrounds. The software environment runs on a Windows 11 64-bit system with an Intel Core i7-12700H CPU and an NVIDIA GeForce RTX 3090 GPU. The programming language is Python 3.10, the deep learning framework is PyTorch 2.1.0, and CUDA 12.1 is used for efficient parallel computing. The specific parameters for model training are as follows: iteration period is set to 200, batch size is set to 16, workers are set to 3, the initial base learning rate is 0.01, and the weight decay coefficient is configured to 0.0005. The input data size is normalized and preprocessed to a uniform resolution of 640×640 pixels. All experiments are conducted under the same hyperparameters without loading any pre-trained weight coefficients.

[0042] Model performance was evaluated using metrics such as precision, recall, average precision, and frames per second (FPS). Ablation experiments were conducted to compare the effects of different model improvement methods and verify the effectiveness of the above modules in improving model performance.

[0043] The improved structure of the YOLOv8 model is as follows: Figure 7 As shown. Figure 8 The actual test results of the power transmission line showed that birds could be identified and their categories were marked, achieving the expected results.

[0044] The beneficial effects of this invention are as follows: (1) Model improvements significantly enhance detection performance: By introducing an improved feature fusion module, a perceptual detection head, and a PSLoss loss function, the model shows significant improvements in accuracy, recall, and average precision across all classes. On the test set, accuracy is improved by 0.8% compared to YOLOv8, recall by 3.4%, mAP by 2.3%, the number of parameters is reduced by 24.9%, and floating-point operations are reduced by 22.4%. This effectively reduces false positives and false negatives and enhances the detection capability for targets of different sizes.

[0045] (2) This invention can effectively extract the features of birds in complex backgrounds and identify more than 200 bird species. It can then be used in conjunction with bird deterrent devices to accurately drive away the target. In addition, the model can be transferred to other bird detection scenarios and reuse the technical framework.

[0046] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0047] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for bird identification on power transmission lines based on deep learning, characterized in that, include: The video from the power transmission line monitoring camera is collected, and the video is split into video frame images by frame, and each video frame image is converted into a pixel color value array; The pixel color value array is input into the backbone feature extraction part of the target detection network to extract a multi-scale feature map. The multi-scale feature maps are input into the improved feature fusion network in the target detection network to perform attention enhancement and multi-scale feature fusion on the multi-scale feature maps, resulting in fused feature maps corresponding to the four scales P2 to P5. The improved feature fusion network includes a Scale Sequence Attention (SSA) module and a Downsampling Feature Fusion (ADFF) module. The SSA module receives the multi-scale feature maps, enhances the weights of bird target-related features, and suppresses redundant background features. The ADFF module receives the features processed by the SSA module at each scale and performs feature fusion between adjacent scales to generate fused feature maps corresponding one-to-one with the detection layers P2 to P5. The fused feature map is input into the perception detection head module, which performs perception enhancement on the fused feature map at each scale and performs bounding box regression and category prediction to obtain candidate bounding boxes and the confidence scores of bird categories corresponding to each candidate bounding box. During the training phase, a training sample set is constructed based on bird images with real annotations. The training sample set is then used as a supervision signal input to the target detection network, and the network parameters are iteratively updated using the PSLoss loss function. After the PSLoss loss function converges, a bird detection model is obtained. The video frame images collected during the process are converted into the pixel color value array and input into the bird detection model. The output is the transmission line bird detection result with category label and location coordinates.

2. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, Converting each of the video frame images into an array of pixel color values ​​includes: Each video frame image is divided into multiple pixels; Each pixel's continuous color value is quantized into an integer between 0 and 255, and the pixel's position in the video frame image and its corresponding color channel value are organized as array elements to form the pixel color value array.

3. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The backbone feature extraction part of the target detection network includes a convolutional structure, a C2f structure, and a spatial pyramid pooling module connected in sequence. When the pixel color value array is input into the backbone feature extraction part of the target detection network, multi-level image features are first extracted through the convolutional structure and the C2f structure, and then the features of different receptive fields are aggregated through the spatial pyramid pooling module to improve the adaptability to bird targets of different sizes.

4. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The Scale Sequence Attention (SSA) module performs global pooling on the input multi-scale feature map, compressing the spatial information of each channel into a global feature value. This feature value is then sequentially transformed and preliminarily filtered through a first fully connected layer and a nonlinear transformation with a ReLU activation function. Next, a second fully connected layer and a nonlinear transformation with a Sigmoid activation function map the feature value to a range of 0 to 1 to generate channel-level attention weights. These channel-level attention weights are then multiplied element-wise with the original features and added to them using an addition operation. This process enhances important channel features relevant to bird target recognition while suppressing redundant background features.

5. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The downsampling feature fusion module ADFF takes large-scale, medium-scale, and small-scale features as input. It performs max pooling and convolution on the large-scale features to achieve downsampling, average pooling and convolution on the medium-scale features to extract medium-scale features, and direct convolution on the small-scale features to extract fine-grained features, thus obtaining three feature maps with the same spatial size. The three feature maps with the same spatial size are then connected along the channel dimension to generate a fusion feature map for identifying densely overlapping bird targets.

6. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The improved feature fusion network also includes an additional P2 detection layer specifically designed for small target features, which is set in the neck structure of the target detection network. The additional P2 detection layer is connected to an improved scale sequence feature fusion structure composed of two downsampled feature fusion modules ADFF and two scale sequence attention modules SSA, so that the fused feature maps corresponding to the four scales P2 to P5 correspond one-to-one with the P2 to P5 detection layers, thereby improving the detection capability of dense small target birds.

7. The deep learning-based bird identification method for power transmission lines according to claim 6, characterized in that, The fusion feature maps corresponding to the four scales P2 to P5 are generated by the additional P2 detection layer. The resolution of the fusion feature map at the P2 scale is 160×160 pixels. It is used to retain the semantic information related to small-scale bird targets under the 160×160 pixel resolution and to achieve accurate detection and localization of the small-scale bird targets. It works in conjunction with the detection layers corresponding to the P3, P4 and P5 scales to improve the detection sensitivity of the small-scale bird targets while maintaining the ability to recognize large-scale bird targets.

8. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The perception detection head module includes a perception module, which is configured to receive the fused feature map from four scales from P2 to P5, and sequentially perform a 1×1 kernel-based group normalization convolution operation and the processing of the perception module on the fused feature map at each scale. The sensing module includes: The first perceptual branch, which consists of average pooling, 1×1 convolution, ReLU activation unit and Sigmoid activation unit in sequence, is used to generate scale attention weights and weight the fused feature maps at different scales. The second perception branch, which consists of deformable convolution, 3×3 convolution and sigmoid activation unit in sequence, is used to generate spatial attention weights and enhance the local spatial features of occluded bird targets and small-scale bird targets. The third perception branch consists of average pooling, a fully connected layer, a ReLU activation unit, a fully connected layer, and a normalization operation Normalize in sequence. The normalization operation Normalize is implemented by a hard sigmoid function to limit the output value to the range of 0 to 1, thereby normalizing the perception features processed by the perception module.

9. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The PSLoss loss function is a total loss function constructed based on the PSIoU index. The PSLoss loss function includes a distribution focus loss term for category prediction, a PSIoU-based bounding box loss term for bounding box regression, and a binary cross-entropy loss term for target presence determination. Specifically, the PSIoU-based bounding box loss term introduces a scaling factor (scale) and a shape cost (Ω) when calculating the overlap between the predicted and ground truth bounding boxes. The scaling factor (scale) is related to the size of the targets in the dataset and is used to control the scale of the auxiliary bounding boxes. The shape cost (Ω) characterizes the shape differences of the bounding boxes and, combined with the width and height of the ground truth bounding boxes, the distance between the center points of the predicted and ground truth bounding boxes, and the positional relationships of the left, right, top, and bottom boundaries of the predicted and ground truth bounding boxes, to improve the accuracy of position regression for small-scale bird targets and accelerate the convergence of the loss function.

10. The deep learning-based bird identification method for power transmission lines according to claim 1, characterized in that, The training sample set includes bird images from the open-source bird image dataset CUB-200 and additional images of common birds. The open-source bird image dataset CUB-200 includes 11,788 images of 200 different bird species. During training, the input data of the training sample set is preprocessed to a uniform resolution of 640×640 pixels. Without loading pre-trained weights, the iteration period is set to 200 times, the batch size is 16, the initial base learning rate is 0.01, and the weight decay coefficient is 0.0005 to improve the robustness of the bird detection model in complex power transmission line scenarios.

Images