Power transmission line bird damage related bird classification method based on image recognition

Abstract

The invention discloses a power transmission line bird damage related bird classification method based on image recognition. The requirements of removing motion blur interference and accurately classifying bird types during power transmission line bird damage related bird classification are met. According to the method, a DebuurGANv2-DRM is combined with a detail recovery module (DRM) to carry out image deblurring processing so as to improve the definition of an input image; image features are extracted through a MobileNetV3 network structure of a depth separable convolution; and in combination with space and channel attention modules, the attention capability of the model on key features is enhanced. The method has the advantages of high classification precision, wide application range, strong motion blur removal capability, lightweight network model, high operation speed and the like, can meet the bird identification requirement under the background of the power transmission line, remarkably reduces the difficulty of bird damage prevention and control of the power transmission line, and has important application value in ecological research and bird damage prevention and control.

Description

Technical Field

[0001] The invention belongs to the technical field of power grid bird damage analysis, and in particular relates to a method for classifying birds related to power transmission line bird damage based on image recognition. Background Art

[0002] With the continuous expansion of the power grid, the continuous improvement of the ecological environment, the continuous increase in the number and range of birds, the contradiction between bird activities and power grid operation has become increasingly prominent, and bird damage to transmission lines has become one of the important reasons for damage to power grid equipment. Bird damage to transmission lines has seriously reduced the stability of power system operation, threatened the safety of the entire power system, and affected the daily life of the public. Therefore, more attention should be paid to bird damage prevention and control.

[0003] At present, there are four basic forms of bird damage to transmission lines, namely bird nests, bird droppings, bird body short circuits, and bird pecking faults. In response to these problems, the power industry has proposed many prevention and control methods, such as artificial bird repellent, bird-proof spikes, bird-proof boxes, and bird-proof baffles. These measures have reduced the occurrence of bird damage to transmission lines to a certain extent and ensured the stable operation of the power system. However, practice has shown that these measures still have certain defects, do not meet the requirements of bird protection, and the prevention and control effects need to be further improved. It is very important to take targeted and differentiated prevention and control measures and correctly identify the bird species that cause transmission line failures.

[0004] Methods for bird species identification include image recognition and bird song recognition. There are usually a lot of background noises around power lines, such as wind, electric current, and the operation of mechanical equipment. These noises will be mixed with the bird songs, affecting the clarity of sound collection, making it difficult for the recognition algorithm to extract effective bird song signals from the noise. Images can clearly capture the appearance feature information of bird feathers, such as color, shape, and patterns, effectively improving the accuracy and reliability of recognition, and can solve the above problems well. At present, there are relatively few studies on bird image recognition related to power line bird damage at home and abroad, and most of the existing research focuses on the field of ecology. Existing methods have high requirements for image quality, but in the power line scenario, bird images often have complex backgrounds and are often accompanied by motion blur. Traditional recognition methods perform poorly in this scenario.

[0005] Therefore, the existing classification methods are difficult to meet the needs of bird classification related to power line bird damage, and there is an urgent need to optimize the existing methods to improve their performance in practical applications. Summary of the invention

[0006] Aiming at the problems of severe motion blur of bird images on power lines, low classification accuracy and high computing power requirements, the present invention discloses a method for classifying birds related to power line bird damage based on image recognition.

[0007] This method processes and classifies images of birds related to power line bird damage obtained by high-definition cameras, and proposes a bird image deblurring method based on DeblurGANv2-DRM. It uses the residual UNet structure combined with the detail recovery module (DRM) to gradually generate clear bird images. Then, a lightweight model for bird image classification based on MobileNetV3 is constructed to classify bird species, and finally achieves accurate recognition of bird species under motion blurred background. This method has the advantages of high classification accuracy, wide application range, effective removal of motion blur interference, high degree of lightweight network model, and fast computing speed.

[0008] A method for classifying birds related to bird damage to power lines based on image recognition comprises: firstly obtaining images of birds related to bird damage to power lines and performing preprocessing; for images with motion blur, deblurring the images using a bird image deblurring method based on DeblurGANv2-DRM to obtain clear bird images; inputting the clear bird images into a constructed lightweight network for bird image classification based on MobileNetV3 to classify the bird species, and outputting recognition results.

[0009] In the lightweight network for bird image classification based on MobileNetV3, the channel attention mechanism and the spatial attention mechanism are introduced to further extract the corresponding spatial area features and channel features in the initial feature map.

[0010] Furthermore, a method for classifying birds related to power line bird damage based on image recognition includes the following steps:

[0011] (1) Obtain images of birds related to power line damage using high-definition cameras;

[0012] (2) Cropping and denoising the bird image data on power lines collected by high-definition cameras to ensure the quality of the image data;

[0013] (3) Perform data augmentation operations such as random cropping, rotation, and horizontal flipping on the preprocessed images to enhance the generalization ability of the model and normalize the images;

[0014] (4) An image deblurring method based on DeblurGANv2-DRM is proposed. This method uses the residual UNet structure, combined with multi-layer convolution, adaptive convolution, local attention mechanism and jump connection, to gradually generate clear bird images and remove the interference of motion blur.

[0015] (5) Build a lightweight network for bird image classification based on MobileNetV3, combine channel and spatial attention mechanisms, extract local and global features of images through deep separable convolution, and classify bird species through fully connected layers;

[0016] (6) Apply the trained neural network to new bird image data and output recognition results to achieve accurate classification of the target bird species.

[0017] The data enhancement operation described in step (3) is only required in the model training stage. When performing actual classification, this operation can be omitted and the preprocessed image can be directly normalized.

[0018] Furthermore, during actual classification, the preprocessing includes:

[0019] (1-1) Crop the region of interest of the bird in the image;

[0020] (1-2) Standardize the size of the cropped image;

[0021] (1-3) Using Gaussian filtering to denoise the image after size standardization;

[0022] (1-4) Normalize the image.

[0023] Furthermore, in the training phase of the model, the preprocessed images are processed to enhance the generalization ability of image data, and the processing methods include one or more of random cropping, random rotation, and horizontal flipping. Through the random cropping operation, different areas are intercepted from the preprocessed images to ensure that the model can adapt to bird images of various sizes and positions during the training process. The data samples are further enriched through random rotation and horizontal flipping. The pixel values ​​of the image are scaled to [-1,1] through normalization processing to ensure the consistency of the image data and accelerate the model training process. The generalization ability and stability of the model in different scenarios are improved, providing more diverse input data for subsequent classification.

[0024] Furthermore, in the bird image deblurring method based on DeblurGANv2-DRM, the feature maps output by the corresponding layers of the encoder and decoder are directly connected by using jump connections to obtain the fused features of the current layer of the decoder; at the same time, the detail recovery module is used in the encoder to further extract the detail information of the fused features of each layer to obtain the refined feature map; finally, the deblurred clear image is output after upsampling.

[0025] Furthermore, in each layer of the encoder, the input feature map is first convolved to achieve downsampling, and then activated by applying the ReLU activation function to obtain the output feature map. In the decoder, the spatial resolution of the feature map is gradually improved by deconvolution, and its input feature map is the result of deconvolution, fusion and detail restoration of the feature map of the previous layer.

[0026] In this method, the feature maps of the corresponding layers of the encoder and decoder are directly connected through skip connections to retain low-level edge and texture information and enhance the reconstruction capability of the decoder.

[0027] In this method, the features fused by the jump connection are input into the detail restoration module (DRM) to further enhance the detail restoration effect. DRM first generates the weight matrix A of the feature map through local attention. The local attention mechanism weights the weight matrix A to ensure that the key areas of the image receive extra attention. Then, the weight map A is used to perform adaptive convolution on the feature map. By adjusting the weight of the feature map, the network can capture the detail information, further remove blur, and generate a refined feature map.

[0028] Finally, the last layer of the decoder is used to upsample the refined feature map to restore the feature map to the resolution of the original image. The skip connection introduces the feature information extracted earlier into this stage, helping to generate clearer images.

[0029] During the training process of DeblurGANv2-DRM, the discriminator D is used to evaluate the difference between the image generated by the model and the real clear image. The deblurred image passes through the discriminator network and is compared with the real clear image to determine its authenticity and perform back propagation control on the network model. In addition to the adversarial loss of GAN, DeblurGANv2-DRM also uses perceptual loss and pixel loss to further optimize the output, making the generated deblurred image visually closer to the real image.

[0030] The MobileNetV3 lightweight network uses an automatically searched and optimized network structure, and reduces the amount of computation in the system by mixing depth-separable convolutions, making the network model more lightweight. However, in order to reduce the amount of computation while ensuring the recognition accuracy of the model, this method adjusts the network structure of MobileNetV3 to further enhance the classification effect. The channel and spatial attention mechanisms are introduced, and the spatial attention module and channel attention module are set to enhance the model's ability to capture key features.

[0031] Furthermore, in the lightweight network for bird image classification based on MobileNetV3:

[0032] (1) First, obtain the preliminary feature map through depth-wise separable convolution;

[0033] (2) These feature maps are then input into the spatial attention module and the channel attention module for processing, respectively, to obtain the corresponding spatial feature maps and channel feature maps;

[0034] (3) Fusing the spatial feature map and the channel feature map to obtain fusion features;

[0035] (4) Upsample the fused features to obtain the final classification result.

[0036] Furthermore, in the channel attention module, the input feature map is subjected to global average pooling and global maximum pooling, the two pooling results are spliced ​​together, and the key features are extracted through a 1×1 convolution layer; the key features are then input into the Sigmoid layer, and the channel attention weights are output; finally, the channel feature map is obtained by multiplying the original input feature map with the channel attention weights.

[0037] Furthermore, in the spatial attention module, global average pooling and global maximum pooling are performed on the original input feature map, and the two pooling results are spliced. The spliced ​​result is then processed through a 1×1 convolution layer and a sigmoid activation function to output the spatial attention weight; finally, the spatial feature map is obtained by multiplying the original feature map with the spatial attention weight.

[0038] This method fuses the feature maps generated by the two attention modules according to their weights. Through this weighted merging of features, it not only retains the richness of semantic information but also reduces unnecessary noise interference.

[0039] In step (4), the classification result can be output by using the original MobileNetV3 network operation for the obtained fusion features, that is, firstly compressing the spatial dimension of the feature map (fusion feature) through the global average pooling operation and converting it into a one-dimensional vector containing global feature information. Subsequently, the one-dimensional vector is input into the fully connected layer and mapped to the classification space after linear transformation to generate a score vector corresponding to the number of classification task categories. These scores are converted into probability distributions through the Softmax function, where the probability value of each category reflects the model's confidence in the category. Finally, the network takes the category with the largest probability value as the final classification result to complete the bird classification task.

[0040] The present invention discloses a method for classifying birds related to bird damage to power lines under a blurred background, and provides a solution to the problems of difficulty in eliminating motion blur in images of bird damage to power lines, low classification accuracy, and high requirements on hardware computing power.

[0041] This method uses DeblurGANv2-DRM combined with a detail restoration module (DRM) to perform image deblurring to improve the clarity of the input image; extracts image features through a MobileNetV3 network structure with deep separable convolution; and combines spatial and channel attention modules to enhance the model's ability to focus on key features. This method has the advantages of high classification accuracy, wide application range, powerful motion blur removal capability, lightweight network model, and fast computing speed. It can meet the needs of bird identification under the background of power transmission lines, significantly reduce the difficulty of bird damage control on power transmission lines, and has important application value in ecological research and bird damage control.

[0042] This method can achieve accurate bird classification under blurred background, and has the advantages of high classification accuracy, wide application range, effective removal of motion blur interference, lightweight network model, fast computing speed, etc. Through the application of the present invention, the accuracy of bird classification related to power line bird damage can be significantly improved, the influence of blurred background can be significantly reduced, and the high requirements of the classification network on hardware can be effectively reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of the steps for classifying birds related to bird damage to power lines in this method. DETAILED DESCRIPTION

[0044] In order to describe the present invention in more detail, the technical solution of the present invention is described in detail below in conjunction with specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0045] like Figure 1 As shown, a method for classifying birds related to power transmission line bird damage based on image recognition includes the following steps:

[0046] (1) Image preprocessing.

[0047] The region of interest (RoI) of the bird in the image is obtained by manual annotation, and the coordinates (x_min, y_min) and (x_max, y_max) of the bounding box are determined, which represent the upper left corner and lower right corner of the bird in the image, respectively. The image is cropped according to these coordinates, retaining only the bird-related area and removing irrelevant background. Next, the cropped image is normalized and resized to 512×512 pixels to ensure that all images have uniform dimensions when input to the classification model.

[0048] After that, the pixels of the cropped image are selected and a 3×3 window is used to generate a Gaussian weight matrix. The pixel values ​​in the window are weighted averaged, the value of the central pixel is replaced, and the entire image is traversed to process the pixels one by one for Gaussian filtering and denoising. Finally, a clear bird image after denoising is obtained, providing higher quality data for subsequent classification.

[0049] (2) The generalization ability of image data is enhanced.

[0050] Through random cropping operations, different regions are captured from the preprocessed images to ensure that the model can adapt to bird images of various sizes and positions during training. Next, random rotation is applied to rotate the image at different angles within the range of ±180 degrees. By horizontal flipping, mirror images of birds are generated to further enrich the data samples. After completing the above data augmentation operations, the image is normalized. The pixel values ​​of the image are scaled to [-1,1] to ensure the consistency of the image data and accelerate the model training process. The generalization ability and stability of the model in different scenarios are improved, providing more diverse input data for subsequent classification.

[0051] When the method of the present invention is used for actual classification, there is no need to perform the operation of step (2). After the Gaussian filtering in step (1), the above-mentioned normalization process can be directly performed.

[0052] (3) Image deblurring processing.

[0053] In order to improve the accuracy of bird image recognition, when the image is motion blurred (which can be identified by human eye recognition), the image needs to be deblurred. Of course, when the bird image is not motion blurred, the subsequent classification operation can be performed directly, omitting this step.

[0054] For images with motion blur, first, the blurred bird image I obtained by the high-definition camera is input into the generator (encoder) of the DeblurGANv2-DRM network. The blurred image is downsampled step by step through multiple convolutional layers to extract the deep feature representation of the image. The ReLU activation function is applied after each convolution operation. The output feature of the image at the encoder layer is F l , then the l+1th layer is the output F l+1 The convolution operation formula is:

[0055] F l+1 =ReLU(Conv(F l ,W l )+b l )

[0056] Among them, W l represents a convolution kernel of size 3*3, b l is the bias term, ReLU is the activation function, and Conv is the convolution operation. This stage extracts the spatial features of the image to capture the structural information of the image.

[0057] The decoder gradually increases the spatial resolution of the feature map through deconvolution, and its input feature map F L-lIt is the result of deconvolution, fusion and detail information capture of the feature map of the previous layer.

[0058] The feature maps of the corresponding layers of the encoder and decoder are directly connected through skip connections to retain low-level edge and texture information and enhance the reconstruction ability of the decoder. l , the Llth layer of the decoder will add it element by element, the specific formula is:

[0059] F′ D,L-l =F D,L-l +F E,l

[0060] Among them, F′ D,L-l is the fusion feature of the current layer of the decoder, F D,L-l is the deconvolution output of the Llth layer of the decoder, F E,l is the output of the encoder layer l.

[0061] The feature F′ after skip connection fusion D,L-l It will be input into the detail restoration module (DRM) to further enhance the detail restoration effect. DRM first generates the weight matrix A of the feature map through local attention. The local attention mechanism weights the weight matrix A to ensure that the key areas of the image receive extra attention. The specific calculation is:

[0062] A=σ(Conv(F′ D,L-l ,W A ))

[0063] Among them, σ represents the sigmoid activation function, W A The convolution kernel is 1*1 in size.

[0064] Then, the weight map A is used to perform adaptive convolution on the feature map. By adjusting the weight of the feature map, the network can capture detailed information and further remove blur. The calculation formula is:

[0065] F″ D,L-l =Conv(A·F′ D,L-l ,W adaptive )

[0066] Where F″ D,L-l is the feature map after the refinement, W adaptive is an adaptive convolution kernel.

[0067] Then the last layer of the decoder upsamples the refined feature map and restores the feature map to the resolution of the original image to obtain a clear image I after deblurring. deblur The skip connection introduces feature information extracted earlier into this stage, helping to generate clearer images.

[0068] During the model training phase, DeblurGANv2-DRM uses the discriminator D to evaluate the difference between the generated image and the real clear image. deblur After passing through the discriminator network, it is compared with the real clear image to determine its authenticity. The confrontation process between the generator and the discriminator is represented by the following loss function L GAN control:

[0069] L GAN =logD(I clear )+log(1-D(I deblur ))

[0070] Among them, I clear It is a real clear image. deblur is the deblurred image output by the generator. The goal of the generator is to minimize this loss function so that D cannot distinguish I deblur and I clear .

[0071] In addition to the adversarial loss of GAN, DeblurGANv2-DRM also uses the perceptual loss L perceptual and pixel loss L pixel To further optimize the output, the generated deblurred image is visually closer to the real image. The perceptual loss measures the difference between the deblurred image and the clear image by comparing their feature representations. The calculation formula is:

[0072]

[0073] Among them, φ i is the feature map of the pre-trained network (generator) at layer i. This loss helps the generated images to be perceptually closer to real images.

[0074] Pixel loss compares the pixels of the blurred image and the clear image to determine the pixel difference between the generated image and the target image. The calculation formula is:

[0075]

[0076] Among them, I i is the i-th pixel value in the real clear image. is the i-th pixel value in the resulting deblurred image. N is the total number of pixels in the image.

[0077] Finally, after this series of processing, the blurred image is deblurred to generate a clear bird image.

[0078] (4) Improve the MobileNetV3 network structure.

[0079] During classification, the deblurred clear image is input into the improved MobileNetV3 network structure, and finally the classification structure is obtained.

[0080] The MobileNetV3 lightweight network uses an automatically searched and optimized network structure, and reduces the amount of computation in the system by mixing depth-separable convolutions, making the network model more lightweight. However, in order to reduce the amount of computation while ensuring the recognition accuracy of the model, this method adjusts the network structure of MobileNetV3 to further enhance the classification effect. The channel and spatial attention mechanisms are introduced, and the spatial attention module and channel attention module are set to enhance the model's ability to capture key features.

[0081] In this process, the model first obtains the preliminary feature map F through the depthwise separable convolution in the original MobileNetV3, and then inputs these feature maps into two attention modules for processing. The spatial attention module uses global average pooling and maximum pooling to capture important spatial features in the image and generate a more comprehensive spatial feature map. The channel attention module performs global average pooling and maximum pooling on the features of each channel, concatenates the pooling results, and uses the activation function to redistribute the weights of the channel features to highlight the more important features in the classification task.

[0082] Next, the feature maps generated by the two attention modules are fused according to their weights. This weighted merging of features not only retains the richness of semantic information, but also reduces unnecessary noise interference. Finally, on the basis of reducing the amount of model calculation, MobileNetV3 successfully achieved accurate classification of bird images and significantly improved the accuracy of recognition.

[0083] The following is a more detailed description of setting up the attention module:

[0084] In order to improve the model's ability to focus on important features, this method introduces a spatial attention module and a channel attention module based on MobileNetV3. It aims to emphasize important spatial regions and channel features in the image through dynamic weighting, thereby enhancing the classification performance of the model. The core idea of ​​the attention module is to assign a weight to each spatial position and each channel. In this way, the model can focus on more important areas and suppress irrelevant background interference and channel information.

[0085] By performing global average pooling and global maximum pooling on the input feature map, the global information and significant features of each channel are obtained. Then, the two pooling results are spliced ​​together and passed through a 1×1 convolution layer to reduce the dimension and extract key features. Subsequently, the feature map processed by the Sigmoid activation function is input to the Sigmoid layer, and the importance weight of each channel is output. Finally, by multiplying the original feature map with the channel attention weight, the model can adaptively adjust the features in the channel dimension.

[0086] In order to introduce the spatial attention mechanism, the original feature map is again subjected to global average pooling and global maximum pooling to obtain the statistical features of spatial information and concatenate them. Subsequently, this concatenation result is also processed by a 1×1 convolution layer and a Sigmoid activation function to output the spatial attention weight. Finally, the original feature map is multiplied by the spatial attention weight to adaptively adjust the features in the spatial dimension, emphasize important areas in the image, and suppress the influence of irrelevant background. By integrating these two attention mechanisms, MobileNetV3 can more effectively extract key information from images, greatly improve the accuracy of bird image classification, and maintain the advantages of lightweight design.

[0087] Perform global average pooling and global maximum pooling operations on the feature map F to obtain channel features:

[0088]

[0089] Where H and W are the height and width of the feature map, respectively. These two operations output a 1×1×C feature vector, representing the global context information.

[0090] F avg With F max Concatenate in the channel dimension:

[0091] F concat =[F avg ,F max ]

[0092] Among them, F concat It is the statistical feature of spatial information after two pooling splicing.

[0093] Perform 1×1 convolution on the concatenated features to reduce the dimension and extract important features:

[0094]

[0095] Among them, W is the convolution kernel, C ′ is the number of output channels.

[0096] The convolved feature map is input into a Sigmoid layer to generate channel attention weights:

[0097] F CA =σ(F conv )

[0098] Among them, σ is the Sigmoid function, and the output weight range is between 0 and 1.

[0099] Multiply the input feature map F by the channel attention weight to enhance important features and obtain the channel feature map

[0100]

[0101] Where ⊙ represents element-by-element multiplication.

[0102] The original feature map is subjected to global average pooling and global maximum pooling again, and after concatenation, it is subjected to 1×1 convolution and Sigmoid activation to generate spatial attention weights:

[0103] F SA =σ(Conv1[F avg ,F max ])

[0104] Multiply with the original feature map, adjust the spatial features, and obtain the spatial feature map

[0105]

[0106] Combine channel and spatial attention features to obtain fusion feature F final :

[0107]

[0108] In generating the final feature map F final Finally, the lightweight bird classification network based on MobileNetV3 first performs a global average pooling operation on the fusion feature map (F final ) and converts it into a one-dimensional vector containing global feature information. Subsequently, the one-dimensional vector is input into the fully connected layer and mapped to the classification space after linear transformation to generate a score vector corresponding to the number of classification task categories. These scores are converted into probability distributions through the Softmax function, where the probability value of each category reflects the model's confidence in the category. Finally, the network completes the bird classification task based on the category with the largest probability value as the final classification result.

[0109] (5) Classify bird images.

[0110] By introducing the residual U-Net structure in DeblurGANv2-DRM, combined with multi-layer convolution, adaptive convolution, local attention mechanism and jump connection, the details of the image are gradually restored and the interference of motion blur is removed, thus generating clearer bird images. The deblurred image will be input into the lightweight bird classification network based on MobileNetV3 for automatic classification of bird species. Finally, the classification result will be displayed in the lower right corner of the original image to help quickly identify the bird species.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the present invention and are not used to limit the scope of protection of the present invention. Even if the present method is described in detail according to the above embodiments, those skilled in the art should understand that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for classifying birds related to power line bird damage based on image recognition, characterized in that: include: The images of birds related to bird damage to power lines are obtained and preprocessed; for images with motion blur, the images are deblurred using a bird image deblurring method based on DeblurGANv2-DRM to obtain clear bird images; the clear bird images are input into a lightweight network for bird image classification based on MobileNetV3 to classify the bird species, and the recognition results are output; in the lightweight network for bird image classification based on MobileNetV3, the channel attention mechanism and the spatial attention mechanism are introduced to further extract the corresponding spatial features and channel features in the initial feature map.

2. The method for classifying birds related to power line bird damage based on image recognition according to claim 1, characterized in that: The pre-processing comprises: (1) Crop the region of interest of the bird in the image; (2) Standardize the size of the cropped image; (3) Using Gaussian filtering to denoise the image after size standardization; (4) Normalize the image.

3. The method for classifying birds related to power line bird damage based on image recognition according to claim 1, characterized in that: During the model training phase, the preprocessed images are processed to enhance the generalization capability of image data. The processing methods include one or more of random cropping, random rotation, and horizontal flipping.

4. The method for classifying birds related to power line bird damage based on image recognition according to claim 1, characterized in that: In the bird image deblurring method based on DeblurGANv2-DRM, the feature maps output by the corresponding layers of the encoder and the decoder are directly connected by using a skip connection to obtain the fusion features of the current layer of the decoder; At the same time, the detail recovery module is used in the encoder to further extract the detail information of each layer of fusion features to obtain the refined feature map, and finally the last deconvolution layer is used to upsample and output the deblurred clear image.

5. The method for classifying birds related to power line bird damage based on image recognition according to claim 4, characterized in that: In each layer of the encoder, a convolution operation is first performed on the input feature map to achieve downsampling, and then a ReLU activation function is applied to activate it to obtain an output feature map.

6. The method for classifying birds related to power line bird damage based on image recognition according to claim 1, characterized in that: In the lightweight network for bird image classification based on MobileNetV3: (1) First, obtain the preliminary feature map through depth-wise separable convolution; (2) These feature maps are then input into the spatial attention module and the channel attention module for processing, respectively, to obtain the corresponding spatial feature maps and channel feature maps; (3) Fusing the spatial feature map and the channel feature map to obtain fusion features; (4) Upsample the fused features to obtain the final classification result.

7. The method for classifying birds related to power line bird damage based on image recognition according to claim 6, characterized in that: In the channel attention module, the input feature map is subjected to global average pooling and global maximum pooling, the two pooling results are concatenated together, and the key features are extracted through a 1×1 convolution layer; the key features are then input into the Sigmoid layer, and the channel attention weights are output; finally, the channel feature map is obtained by multiplying the original input feature map with the channel attention weights.

8. The method for classifying birds related to power line bird damage based on image recognition according to claim 6, characterized in that: In the spatial attention module, global average pooling and global maximum pooling are performed on the original input feature map, and the two pooling results are spliced. The spliced ​​result is then processed through a 1×1 convolution layer and a Sigmoid activation function to output the spatial attention weight. Finally, the spatial feature map is obtained by multiplying the original feature map with the spatial attention weight.

9. The method for classifying birds related to power line bird damage based on image recognition according to claim 1, characterized in that: The classification results will be displayed on the original image.

Images