Bird small target detection and positioning method and system under position awareness and frequency enhancement

Abstract

The present application relates to the technical field of image recognition, and discloses a bird small target detection and positioning method and system under position sensing and frequency enhancement, comprising: obtaining a substation bird small target image dataset, which is a binocular image dataset; inputting a left eye image in the binocular image into a neural network to obtain position information of a bird small target in the left eye image; inputting the binocular image and the position information of the bird small target in the left eye image into a backbone feature extraction network to extract multi-scale features, and constructing a multi-level cost volume based on the extracted multi-scale features; performing parallax iterative update calculation on the constructed multi-level cost volume based on a multi-frequency channel attention mechanism, optimizing the parallax result through a loss function, and realizing spatial positioning of the bird small target. The present application can realize efficient detection and accurate positioning.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method and system for detecting and locating small bird targets using location awareness and frequency enhancement. Background Technology

[0002] As the core hub of the power system, the safe and stable operation of substations directly affects the reliability of power supply. When birds nest, roost, or fly in substation areas, they are highly susceptible to collisions with power equipment such as utility poles and insulator strings. Their behavior, such as excreting waste or building nests, can also cause short circuits and flashovers, seriously threatening power grid safety and potentially causing significant economic losses. Therefore, accurate detection and three-dimensional spatial positioning of small bird targets in substation environments are crucial prerequisites for targeted bird removal and prevention of power safety risks. This provides clear target location information for subsequent operations such as laser-based bird removal, ensuring the efficiency and accuracy of these operations.

[0003] However, existing technologies face many challenges in bird detection and location tasks at substations, mainly including: 1. In the detection of small bird targets, the complex environment of substations presents multiple challenges to the detection task. First, small bird targets are inherently small in size, occupying a very low percentage of pixels in the image, resulting in a lack of main texture information and blurred edge features, making it difficult for traditional target detection networks to effectively extract discriminative features. Second, the dense distribution of objects such as power poles, insulator strings, trees, and vegetation within substations not only easily obscures birds partially or completely, but some objects also share visual features with birds, easily leading to missed or false detections by the algorithm. Third, substations are mostly located in open outdoor environments, where factors such as changes in light intensity, extreme weather interference, and smoke and dust obstruction frequently occur, further reducing the stability of the detection algorithm and failing to provide reliable target area references for subsequent localization tasks.

[0004] 2. In the 3D spatial localization of birds, the core technology relies on stereo matching to obtain disparity information, and then calculates the target depth and spatial coordinates. However, existing stereo matching methods have significant shortcomings. Traditional stereo matching methods require manual extraction of feature points before searching and matching, which is not only cumbersome and slow, but also has a high false matching rate, making it difficult to meet the needs of real-time detection and localization. Although deep learning stereo matching methods developed in recent years have achieved end-to-end learning and simplified the operation process, they still perform poorly when dealing with occluded areas, weak texture areas, and bird edge details commonly found in substation scenarios. Furthermore, they are prone to losing high-frequency information during disparity iteration updates, resulting in insufficient disparity prediction accuracy. This makes it impossible to provide accurate spatial position parameters for laser deterrence systems, seriously affecting the accuracy of deterrence operations.

[0005] In summary, existing methods for detecting small birds in complex substation scenarios suffer from low accuracy and inaccurate positioning parameters, making it impossible to achieve efficient detection and precise positioning. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a method and system for detecting and locating small bird targets under position awareness and frequency enhancement, which can achieve efficient detection and accurate positioning.

[0007] To address the aforementioned technical problems, this invention provides a method for detecting and locating small bird targets using location awareness and frequency enhancement, comprising: Acquire a dataset of small bird images in substations, wherein the dataset of small bird images in substations is a binocular image dataset; The left eye image from the binocular images is input into the neural network to obtain the position information of the small bird target in the left eye image; The location information of small bird targets in the binocular and left eye images are fed into the backbone feature extraction network to extract multi-scale features, and a multi-level cost body is constructed from the extracted multi-scale features. Based on a multi-frequency channel attention mechanism, the disparity is iteratively updated and calculated using a multi-level cost volume. The disparity results are then optimized under supervision using a loss function, enabling spatial localization of small bird targets.

[0008] Furthermore, the step of inputting the left eye image from the binocular images into the neural network to obtain the positional information of the small bird target in the left eye image includes: Let the left eye image be... Extracting using the YOLOv10 backbone network The multi-layered features are denoted as , , The extraction process includes: The feature map is obtained after processing by two layers of CBS modules, denoted as... ,Will The feature map is obtained by feeding it into the C2f module, which is an improvement based on the multi-scale edge enhancement module, and denoted as . , The first layer of features, denoted as […], is obtained after processing by the CBS module and the C2f module improved based on the multi-scale edge enhancement module. , After spatial channel downsampling, the data is fed into a C2f module improved based on a multi-scale edge enhancement module to obtain the second layer of features, denoted as... , After spatial channel downsampling, the data is fed into a C2f module improved based on a multi-scale edge enhancement module, and then spatial pyramid pooling is used to obtain the feature map, denoted as... , The third layer of features is obtained through the large-kernel spatial attention module, denoted as... ; Will , , The features are fed into a feature pyramid network based on spatial feature reconstruction for feature fusion. The fusion process includes: separating the context in different directions through adaptive horizontal pooling and vertical pooling operations to obtain features of the rectangular region, resulting in two feature maps of different sizes, denoted as . and ;Will and Adding them together yields the feature map after remodeling the rectangular region, denoted as . Then, a weight map is generated using the Sigmoid activation function, denoted as . ;Will and Element-wise multiplication yields the output of the rectangular self-calibration module, denoted as . ;against and Use and obtain The same method is used to obtain the output of the rectangular self-calibration module, denoted as... and ; By adaptive average pooling , , Feature maps adjusted to the same size are denoted as , , ;right , , The result of feature concatenation and enhancement is denoted as: Then Perform channel-dimensional segmentation, and denote the output enhanced features as follows: , , ;Will and The data is fed into the multi-scale feature fusion module, and the fused features are denoted as follows: ;Will and The fusion is performed in the dynamic interpolation fusion module, and the fused features are denoted as follows. ;Will and After fusion by the multi-scale feature fusion module, the fused features are denoted as follows: ;Will and The fusion is performed in the dynamic interpolation fusion module, and the fused features are denoted as follows. ;Will and The feature output is calculated by the multi-scale feature fusion module, and the output feature is denoted as . ; Will , , The data is fed into the detection head, and the location information of the small bird target is calculated.

[0009] Furthermore, the aforementioned The feature map is obtained by feeding it into the C2f module, which is an improvement based on the multi-scale edge enhancement module, and denoted as . ,include: First, the size is gradually reduced by using the initial convolutional layers and spatial channel downsampling of the YOLOv10 backbone network, while extracting basic features. The resulting feature map is denoted as... Then on Adaptive average pooling is performed, the number of channels is reduced through convolution operations, and then spatial details are extracted through standard convolution operations to obtain multi-scale features, denoted as . ; Average pooling is used for multi-scale features Smoothing is performed to generate low-frequency features, denoted as... ;right and Performing difference operations preserves high-frequency information in rapidly changing regions of the image, yielding high-frequency features, denoted as... ; Channel information is fused through convolution operations, and a high-frequency feature weight map is generated using the Sigmoid activation function, which is then combined with... Superposition yields the output features enhanced with high-frequency information, denoted as... ; Feature maps processed by the edge enhancement module at different scales Bilinear interpolation is performed to upsample to the input size to maintain spatial alignment, resulting in the output of each branch, denoted as... For multi-scale features The local features are obtained through convolution operations and denoted as follows: ; Will The enhanced features are concatenated with the outputs of each branch along the channel dimension to obtain the enhanced features, denoted as . ; Next to Perform average pooling and max pooling operations to obtain the channel-compressed feature map, denoted as . ; Depth-separable convolution pairs Deep feature extraction is performed, and residual connections are used to avoid excessive suppression of original features by spatial attention. The resulting features are denoted as follows: ; Will The low-frequency components obtained through global average pooling are denoted as follows: and with Perform a difference operation to obtain the high-frequency component, denoted as . Next, and Residual modulation is performed to obtain an enhanced feature map, denoted as... ; Will The number of channels is adjusted through convolution operations to obtain the final output features. .

[0010] Furthermore, the aforementioned The third layer of features is obtained through the large-kernel spatial attention module, specifically: Extracting from convolutional branches with three different receptive fields Feature maps with different receptive field ranges, denoted as , , ; right , , The correlation between channels was modeled using convolution operations to obtain... ; Connect along the channel dimension using matrix concatenation operations. The fused feature map is obtained and denoted as , ; Again Perform average pooling and max pooling operations to obtain spatial statistics, denoted as... ; Split along channel dimension We obtain three attention weights, denoted as... ; The generated attention weights are multiplied element-wise with the corresponding branch features to obtain the fused features, which are then restored through convolution operations and summed. Multiplying each element together yields the third layer of features.

[0011] Furthermore, the aforementioned and The data is fed into the multi-scale feature fusion module, including: for First, features are extracted using convolutional layers. Then, the output range of the features is adjusted using the sigmoid activation function. The adjusted feature map is denoted as... ;Will and Perform element-wise multiplication to obtain ; The and The fusion is performed in the dynamic interpolation fusion module, including: firstly, ... Perform interpolation upsampling and adjust its size to match... The same, after further convolution operations and Integration, to obtain .

[0012] Furthermore, the step of feeding the positional information of small bird targets in both the binocular and left-eye images into the backbone feature extraction network to extract multi-scale features includes: For the left eye image in a binocular image, the generated multi-scale feature map is denoted as... ; Combine the location information of small bird targets in the left eye image with The resulting features are denoted as follows: ; Will The decoder outputs an enhanced multi-scale feature map, denoted as... ; Used Using the same method, for the right eye image in a binocular image, an enhanced multi-scale feature map is obtained, denoted as... .

[0013] Furthermore, the construction of a multi-level cost body for the extracted multi-scale features includes: Select and Construct an initial cost body, use the dot product of feature vectors as a similarity measure, and calculate... Each pixel and The matching cost for all pixels in the corresponding row is used to construct a 4D correlation cost volume; the 4D cost volume is then converted into a 3D correlation cost volume by calculating the feature dot product under the same y coordinate only. Then, three one-dimensional average pooling operations are performed on the 3D related cost volume to obtain a four-layer multi-level cost volume as the final multi-level cost volume.

[0014] Furthermore, the disparity iterative update calculation for the constructed multi-level cost volume based on the multi-frequency channel attention mechanism includes: Will Contextual features are generated using the ReLU activation function, denoted as […]. Set initial parallax Initiate the parallax iteration process; for First, preprocessing is performed using convolutional layers to obtain... High- and low-frequency separation is achieved by converting contextual features to the frequency domain through two-dimensional discrete cosine transform. Next, the Top-K algorithm is used to select the top K key components with the greatest impact on matching accuracy from all frequency domain components based on a predefined frequency index. Redundant low-frequency noise and invalid high-frequency interference are eliminated, balancing feature representation and computational efficiency. Statistical information is extracted from the selected K frequency feature maps using global average pooling, global max pooling, and global min pooling, respectively, resulting in three statistical vectors, denoted as . The three statistical vectors are first passed through a dimension-reducing fully connected layer, then processed by the ReLU activation function, then through a dimension-upgrading fully connected layer to restore the number of channels, and finally the output is mapped to the [0,1] interval by the Sigmoid activation function to generate frequency attention weights, denoted as . ; Will The hidden information of the gated recurrent unit at the corresponding level is multiplied element by element to achieve adaptive enhancement of information of different frequencies, and the hidden information of the gated recurrent unit after correction by the multi-frequency channel attention mechanism is obtained.

[0015] Furthermore, the step of supervising and optimizing the disparity results through a loss function to achieve spatial localization of small bird targets includes: The L1 loss function is used to supervise the disparity iteration process, and the network parameters are optimized through backpropagation to reduce disparity error. The optimized disparity results are upsampled to restore the original image resolution, resulting in the final disparity map; The intrinsic and extrinsic parameters of the binocular camera are calculated, the depth information of small bird targets is calculated based on the principle of triangulation, and then the three-dimensional spatial coordinates of the target are obtained through coordinate transformation to achieve precise positioning.

[0016] The present invention also provides a small bird target detection and localization system based on position awareness and frequency enhancement, comprising: The data acquisition module is used to acquire a dataset of small bird target images in substations, wherein the dataset of small bird target images in substations is a binocular image dataset; The bird target location information acquisition module is used to input the left eye image from the binocular images into the neural network to obtain the location information of bird targets in the left eye image; The multi-scale feature extraction module is used to input the location information of small bird targets in the binocular image and the left eye image into the backbone feature extraction network to extract multi-scale features. The multi-level cost body construction module is used to construct a multi-level cost body from the extracted multi-scale features. The disparity iterative update module is used to perform disparity iterative update calculations on the constructed multi-level cost volume based on a multi-frequency channel attention mechanism. The bird small target localization module is used to achieve spatial localization of bird small targets by supervising and optimizing the disparity results through a loss function.

[0017] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention enhances the edge details of small targets, captures long-range dependencies, and aggregates multi-scale contextual information. It also improves the quality of input images by combining image super-resolution reconstruction technology, enabling the network to extract features of small bird targets in complex substation scenarios more fully. This effectively reduces false negatives and false positives caused by factors such as low pixel ratio of small targets, heavy background interference, and changes in illumination. At the same time, the synergy of position awareness and frequency enhancement clarifies the target region for disparity calculation, avoiding interference from redundant background features, thereby achieving efficient detection and accurate localization. Attached Figure Description

[0018] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart of a method in a preferred embodiment of the present invention.

[0019] Figure 2 This is a diagram showing the overall architecture of the improved YOLOv10 network in a preferred embodiment of the present invention.

[0020] Figure 3 This is a diagram illustrating the process of the backbone feature extraction network extracting multi-scale features in a preferred embodiment of the present invention.

[0021] Figure 4 This is a diagram showing the overall architecture of the disparity calculation network in a preferred embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0023] Reference Figure 1 As shown, this invention discloses a method for detecting and locating small bird targets under location awareness and frequency enhancement, comprising the following steps: S1: Obtain a dataset of small bird images in substations. The dataset of small bird images in substations is a binocular image dataset. In this embodiment, a binocular camera is used to collect images of small birds in substations and construct the dataset of small bird images in substations.

[0024] S1-1: Deploy binocular cameras in different areas of the substation to cover typical interference scenarios such as power poles, insulator strings, and trees, and collect bird images under different lighting and weather conditions to ensure data diversity.

[0025] S1-2: Use Lableme software to manually annotate small bird targets, generate bounding box truth labels, and unify the annotation category as "birds".

[0026] S1-3: Combine the image information obtained from S1-1 and S1-2 to obtain a stereo image dataset.

[0027] S2: Input the left eye image from the binocular images into the neural network to obtain the position information of the small bird target in the left eye image.

[0028] S2-1: Let the left eye image be... Extracting using the YOLOv10 backbone network The multi-layered features are denoted as , , .

[0029] S2-1-1: The feature map is obtained after processing by a two-layer CBS module (the CBS module contains a 3×3 convolutional module Conv, a batch normalization module, and a SiLU activation function module), denoted as . .

[0030] S2-1-2: Will The feature map is obtained by feeding it into the C2f module, which is an improvement based on the multi-scale edge enhancement module, and denoted as . In this embodiment, processing The following example illustrates the processing steps of the C2f module improved based on the multi-scale edge enhancement module.

[0031] S2-1-2-1: First, the size is gradually reduced by using the initial convolutional layers and spatial channel downsampling of the YOLOv10 backbone network, while extracting basic features to generate a feature map of size H×W×C, denoted as H×W×C. Then on Adaptive average pooling is performed, and the number of channels is reduced through a 1×1 convolution operation (in this embodiment, the number of channels is reduced from the original C to C / 4). Then, a 3×3 standard convolution operation is used to extract spatial details, resulting in multi-scale features, denoted as . The calculation formula for this process is: , In the formula, This represents an adaptive average pooling operation, where i = 1, 2, ..., N, and N is the number of scales. In this embodiment, N = 4, corresponding to 4 scale features. ~ .

[0032] S2-1-2-2: Using 3×3 average pooling for multi-scale features Smoothing is performed to generate low-frequency features, denoted as... ;right and Performing difference operations preserves high-frequency information in rapidly changing regions of the image, yielding high-frequency features, denoted as... Filter out areas in the background that change slowly.

[0033] S2-1-2-3: Due to Some channels may have weak responses. Channel information is fused using a 1×1 convolution operation, and a high-frequency feature weight map is generated using the Sigmoid activation function, then combined with… Superposition yields the output features enhanced with high-frequency information, denoted as... The calculation formula is: , In the formula, This represents the Sigmoid activation function.

[0034] S2-1-2-4: Analyzing feature maps processed by the edge enhancement module at different scales. Bilinear interpolation is performed to upsample to the input size to maintain spatial alignment, resulting in the output of each branch, denoted as... Besides multi-scale branching, for multi-scale features Then, a 3×3 convolution operation is performed to obtain local features, denoted as... ;Will The enhanced features are concatenated with the outputs of each branch along the channel dimension to obtain the enhanced features, denoted as . The calculation formula is: .

[0035] S2-1-2-5: Next to... Perform average pooling and max pooling operations to obtain the channel-compressed feature map, denoted as . The calculation formula is: , In the formula, This indicates the average pooling operation. This indicates a max pooling operation.

[0036] S2-1-2-6: Using depthwise separable convolution pairs Deep feature extraction is performed, and residual connections are used to avoid excessive suppression of original features by spatial attention. The resulting features are denoted as follows: The calculation formula is: , In the formula, This indicates that the size of a cascaded set is and Depth-separable convolution, This indicates that the size of a cascaded set is Depth-separable convolution, This indicates that the feature map is copied and expanded along the channel dimension.

[0037] S2-1-2-7: Will The low-frequency components obtained through global average pooling are denoted as follows: and with Perform a difference operation to obtain the high-frequency component, denoted as . Next, and Residual modulation is performed to obtain an enhanced feature map, denoted as... The calculation formula is: , In the formula, This indicates element-wise multiplication.

[0038] S2-1-2-8: Will The final output features are obtained by adjusting the number of channels through a 1×1 convolution operation. ,Right now: .

[0039] S2-1-3: The first layer of features, denoted as […], is obtained after processing by the CBS module and the C2f module improved based on the multi-scale edge enhancement module. .

[0040] S2-1-4: After spatial channel downsampling, the data is fed into a C2f module improved based on a multi-scale edge enhancement module to obtain the second layer of features, denoted as... .

[0041] S2-1-5: After spatial channel downsampling, the data is fed into a C2f module improved based on a multi-scale edge enhancement module, and then spatial pyramid pooling is used to obtain the feature map, denoted as... .

[0042] S2-1-6: The third layer of features is obtained through the large-kernel spatial attention module, denoted as... .

[0043] S2-1-6-1: Receive the feature map after it has been enhanced by the multi-scale edge enhancement module and then spatial pyramid pooling. To adapt to the scale differences of small bird targets in substations, convolutional branches of three different receptive fields were used to extract... Feature maps with different receptive field ranges, denoted as , , The calculation method is as follows: , , ; In the formula, For depthwise separable convolution, For large separable convolutions, K×K represents the convolution kernel.

[0044] S2-1-6-2: Because feature branches with different receptive fields contain different information, it is necessary to generate targeted attention weights through channel modeling for three feature maps with different receptive field ranges. , , The correlation between channels was modeled using 1×1 convolution operations to obtain... The calculation formula is: , , .

[0045] S2-1-6-3: Connecting along the channel dimension using matrix concatenation operations The fused feature map is obtained and denoted as , The calculation formula is: , In the formula, Concat represents the concatenation operation.

[0046] S2-1-6-4: Again Perform average pooling and max pooling operations to obtain spatial statistics, denoted as... The calculation formula is: , In the formula, AP represents average pooling operation and MP represents max pooling operation.

[0047] S2-1-6-5: Splitting along the channel dimension We obtain 3 attention weights, denoted as The calculation formula is: , In the formula, The sigmoid activation function is used, and Split represents the splitting operation.

[0048] S2-1-6-6: The generated attention weights are multiplied element-wise with the corresponding branch features to obtain the fused features, and then the number of channels is restored through a 1×1 convolution operation. Element-wise multiplication yields the final enhanced feature, i.e., the third-layer feature, calculated using the following formula: .

[0049] S2-2: Will , , The features are fed into a feature pyramid network based on spatial feature reconstruction for feature fusion. It receives feature maps from three different levels output by the improved YOLOv10 backbone network. , , , The resolution is 1 / 8 of the original image, which belongs to low-level features. It retains details such as the edges and textures of small bird targets, but the semantic expression is weak. Its resolution is 1 / 16 of the original image, belonging to the mid-level features, balancing details and semantic information; The resolution is 1 / 32 of the original image, representing high-level features with strong semantic expression, but lacking spatial detail. The overall architecture diagram of the improved YOLOv10 network is shown below. Figure 2 As shown.

[0050] S2-2-1: First, for the input features ,by For example, targeting The self-calibrating matrix module enables the network to focus on the features of the target region. First, adaptive horizontal and vertical pooling operations are used to separate the context in different directions to obtain the features of the rectangular region, resulting in two features of different sizes. and Feature map, denoted as and ;Will and Adding them together yields the feature map after remodeling the rectangular region, denoted as . Then, a weight map is generated using the Sigmoid activation function, denoted as . The calculation formula is: , In the formula, , These represent the kernel size as follows: and Large kernel strip convolution, It is the Sigmoid activation function. This represents the ReLU activation function with a batch normalization layer.

[0051] S2-2-2: Will and Element-wise multiplication yields the output of the rectangular self-calibration module, denoted as . ;against and Use and obtain The same method is used to obtain the output of the rectangular self-calibration module, denoted as... and .

[0052] At the same time, adaptive average pooling will be used to... , , Feature maps adjusted to the same size are denoted as , , ;right , , The result of feature concatenation and enhancement is denoted as: The calculation formula is: , In the formula, It is a rectangular self-calibration module with 3 stacked layers.

[0053] S2-2-3: Then... Perform channel-dimensional segmentation, and denote the output enhanced features as follows: , , .

[0054] S2-2-4: Will and The data is fed into the multi-scale feature fusion module, and the fused features are denoted as follows: ;Will and The fusion is performed in the dynamic interpolation fusion module, and the fused features are denoted as follows. ;Will and After fusion by the multi-scale feature fusion module, the fused features are denoted as follows: ;Will and The fusion is performed in the dynamic interpolation fusion module, and the fused features are denoted as follows. ;Will and The feature output is calculated by the multi-scale feature fusion module, and the output feature is denoted as . .

[0055] In this embodiment, fusion and To illustrate the operation process of the multi-scale feature fusion module, the following example will be used: and The data is fed into the multi-scale feature fusion module, including: for First, through a size of Feature extraction is performed using convolutional layers, followed by adjustment of the output range of the features using the sigmoid activation function. The adjusted feature map is denoted as... ;Will and Element-wise multiplication is performed to obtain the final weighted fused feature map, i.e. The calculation formula is: .

[0056] In this embodiment, fusion and To illustrate the operation process of the dynamic interpolation fusion module, the following example will be used: and The fusion is performed in the dynamic interpolation fusion module, including: first, processing the low-resolution feature map... Perform interpolation upsampling to resize the feature map to match the high-resolution feature map. Similarly, after another convolution operation, and... Integration, to obtain The calculation formula is: , In the formula, This is for bilinear interpolation upsampling.

[0057] S2-3: Will , , The data is fed into the detection head, and the location information of the small bird target is calculated.

[0058] The detection head receives three enhanced feature maps output by a feature pyramid network based on spatial feature reconstruction. , , ,in Focus on the details of the target's edges Balancing details and semantics To enhance global semantics, these three features are fed into the detection head, outputting the location information of small bird targets.

[0059] S3: As Figure 3 As shown, the location information of small bird targets in the binocular and left-eye images are fed into the backbone feature extraction network to extract multi-scale features.

[0060] S3-1: Perform layer-by-layer feature extraction on the binocular images. Taking the left eye image as an example, for the left eye image in the binocular images, generate multi-scale feature maps with resolutions of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image resolution, denoted as... Then, the location information of the small bird targets in the left eye image calculated in step S2 is compared with... The resulting features are denoted as follows: This process is implemented through an encoder, specifically as follows: Perform initial convolution on the input image to output a shallow feature map. Preserve the basic contour and texture information of the image; transform the shallow feature map By incorporating the inverted bottleneck module, the number of channels is increased from 64 to 128, and the resolution is downsampled to 1 / 4 using a convolution with a stride of 2, outputting a mid-level feature map. Strengthen the initial distinction between target and background; enhance the mid-layer feature map An additional deep convolutional module is used, which keeps the number of channels unchanged at 128, while the resolution is reduced to 1 / 8 by downsampling with a stride of 2, outputting a mid-to-deep feature map. Highlighting local details of the target; mid-to-deep feature maps By inverting the bottleneck module and using next-generation convolutional networks Network The module reduces the resolution to 1 / 16 by downsampling with a step size of 2, increases the number of channels to 256, and outputs deep feature maps. It possesses strong semantic discrimination capabilities; it integrates deep feature maps. By using an additional deep convolutional module and an inverted bottleneck module, the number of channels remains unchanged at 256, while the resolution is reduced to 1 / 32 through downsampling with a stride of 2, outputting a global feature map. This allows for the global differentiation of targets from complex backgrounds. Then, the feature map... Output features are obtained by fusing with location information. .

[0061] S3-2: Will The decoder outputs an enhanced multi-scale feature map, denoted as... This process is implemented through a decoder, specifically as follows: By constructing a cross-scale fusion module, the features fused with location information are... Corresponding high-level enhancement features To perform feature fusion, firstly... Upsampling via bilinear interpolation With features Alignment is then performed, followed by channel concatenation and 3×3 convolution to achieve interactive fusion of high and low layer features. At the same time, residual connections are used to preserve the original feature information, resulting in a low-level enhanced feature map. ,by For example, after fusing with location information, we get .Will and The samples are fed into a cross-scale fusion module for fusion to obtain... The calculation formula is: , Includes and .

[0062] This provides precise feature support for the construction of the multi-level cost volume in the subsequent step S4, helping to improve the accuracy of disparity calculation and spatial positioning.

[0063] Used Using the same method, for the right eye image in a binocular image, an enhanced multi-scale feature map is obtained, denoted as... .

[0064] S4: Construct a multi-level cost body for the extracted multi-scale features.

[0065] S4-1: Select the left eye feature map at 1 / 4 resolution output from S3. and right eye feature map Construct an initial cost body, use the dot product of feature vectors as a similarity measure, and calculate the left eye feature map. Each pixel and the right eye feature map For each row, the matching cost of all pixels is used to construct a 4D correlation cost volume. The 4D cost volume is then converted to a 3D correlation cost volume by calculating only the feature dot product at the same y-coordinate. The calculation formula is as follows: , In the formula, the subscript i represents the x-coordinate and j represents the y-coordinate.

[0066] S4-2: Then, perform three one-dimensional average pooling operations on the 3D related cost volume to obtain a four-layer multi-level cost volume as the final multi-level cost volume. The four-layer multi-level cost volume can cover matching information at different scales, improving the robustness of subsequent disparity iterations.

[0067] S5: The disparity is iteratively updated and calculated on the constructed multi-level cost volume based on a multi-frequency channel attention mechanism. The overall architecture diagram of the disparity calculation network is shown below. Figure 4 As shown.

[0068] S5-1: The enhanced left eye image's multi-scale features, i.e. Contextual features are generated using the ReLU activation function, denoted as […]. Hidden information used to initialize the gated loop unit Set initial parallax Initiate the parallax iteration process;

[0069] S5-2: Contextual features obtained by the feature extraction network First, it is preprocessed using a 1×1 convolutional layer to obtain... High- and low-frequency separation is achieved by converting contextual features to the frequency domain through two-dimensional discrete cosine transform. Next, the Top-K algorithm is used to select the top K key components with the greatest impact on matching accuracy from all frequency domain components based on a predefined frequency index. Redundant low-frequency noise and invalid high-frequency interference are eliminated, balancing feature representation and computational efficiency. Statistical information is extracted from the selected K frequency feature maps using global average pooling, global max pooling, and global min pooling, respectively, resulting in three statistical vectors, denoted as . The three statistical vectors are first passed through a dimension reduction fully connected layer. Then it is processed by the ReLU activation function, and then by a fully connected layer of increased dimensionality. The number of channels is restored, and finally, the output is mapped to the [0,1] interval using the Sigmoid activation function to generate frequency attention weights, denoted as... The calculation formula is: , In the formula, This represents the Sigmoid activation function. Represents the ReLU activation function; The generated frequency attention weights The hidden information of the gated recurrent unit at the corresponding level is multiplied element-wise to achieve adaptive enhancement of information at different frequencies, resulting in the gated recurrent unit hidden information corrected by the multi-frequency channel attention mechanism. The update formula is as follows: , In the formula, This represents the hidden information of the corresponding level of gated loop unit. This represents the hidden information of the gated loop unit after correction by the multi-frequency channel attention mechanism.

[0070] S6: Spatial localization of small bird targets is achieved by supervising and optimizing the disparity results through the loss function.

[0071] S6-1: The L1 loss function is used to supervise the disparity iteration process. The network parameters are optimized through backpropagation to reduce disparity error. The L1 loss function is as follows: , In the formula, For L1 loss function, For true parallax, To predict parallax, =0.9, This represents the number of iterations.

[0072] S6-2: Upsample the optimized disparity results to restore them to the original image resolution, and obtain the final disparity map;

[0073] S6-3: Calculate the intrinsic and extrinsic parameters of the binocular camera, calculate the depth information of small bird targets based on the principle of triangulation, and then obtain the three-dimensional spatial coordinates of the target through coordinate transformation to achieve precise positioning.

[0074] This invention also discloses a small bird target detection and localization system based on position awareness and frequency enhancement, comprising: The data acquisition module is used to acquire a dataset of small bird target images in substations, wherein the dataset of small bird target images in substations is a binocular image dataset; The bird target location information acquisition module is used to input the left eye image from the binocular images into the neural network to obtain the location information of bird targets in the left eye image; The multi-scale feature extraction module is used to input the location information of small bird targets in the binocular image and the left eye image into the backbone feature extraction network to extract multi-scale features. The multi-level cost body construction module is used to construct a multi-level cost body from the extracted multi-scale features. The disparity iterative update module is used to perform disparity iterative update calculations on the constructed multi-level cost volume based on a multi-frequency channel attention mechanism. The bird small target localization module is used to achieve spatial localization of bird small targets by supervising and optimizing the disparity results through a loss function.

[0075] The present invention also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for detecting and locating small bird targets under position awareness and frequency enhancement.

[0076] The present invention also discloses a device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for detecting and locating small bird targets under position awareness and frequency enhancement.

[0077] Compared with the prior art, the advantages of the present invention are: 1. This invention fully utilizes the synergistic advantages of the multi-scale edge enhancement module, the large-kernel spatial attention module, and the feature pyramid network based on spatial feature reconstruction—enhancing the edge details of small targets, capturing long-range dependencies, and aggregating multi-scale contextual information. At the same time, it combines image super-resolution reconstruction technology to improve the quality of input images, enabling the network to extract features of small bird targets in complex substation scenarios more fully, effectively reducing missed detections and false detections caused by factors such as low pixel ratio of small targets, high background interference, and changes in illumination.

[0078] 2. This invention forms a closed-loop support detection module for accurate identification and reliable positioning by combining position perception and frequency enhancement. This module clarifies the target area for disparity calculation and avoids interference from redundant background features, thereby achieving efficient detection and accurate positioning. It can provide accurate spatial position parameters and provide efficient and reliable full-process technical support for bird removal tasks in substations.

[0079] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0080] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0081] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0082] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0083] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for detecting and locating small bird targets using location awareness and frequency enhancement, characterized in that, include: Acquire a dataset of small bird images in substations, wherein the dataset of small bird images in substations is a binocular image dataset; The left eye image from the binocular images is input into the neural network to obtain the position information of the small bird target in the left eye image; The location information of small bird targets in the binocular and left eye images are fed into the backbone feature extraction network to extract multi-scale features, and a multi-level cost body is constructed from the extracted multi-scale features. Based on a multi-frequency channel attention mechanism, the disparity is iteratively updated and calculated using a multi-level cost volume. The disparity results are then optimized under supervision using a loss function, enabling spatial localization of small bird targets.

2. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 1, characterized in that: The step of inputting the left eye image from the binocular images into the neural network to obtain the position information of the small bird target in the left eye image includes: Let the left eye image be... Extracting using the YOLOv10 backbone network The multi-layered features are denoted as , , The extraction process includes: The feature map is obtained after processing by two layers of CBS modules, denoted as... ,Will The feature map is obtained by feeding it into the C2f module, which is an improvement based on the multi-scale edge enhancement module, and denoted as . , The first layer of features, denoted as […], is obtained after processing by the CBS module and the C2f module improved based on the multi-scale edge enhancement module. , After spatial channel downsampling, the data is fed into a C2f module improved based on a multi-scale edge enhancement module to obtain the second layer of features, denoted as... , After spatial channel downsampling, the data is fed into a C2f module improved based on a multi-scale edge enhancement module, and then spatial pyramid pooling is used to obtain the feature map, denoted as... , The third layer of features is obtained through the large-kernel spatial attention module, denoted as... ; Will , , The features are fed into a feature pyramid network based on spatial feature reconstruction for feature fusion. The fusion process includes: separating the context in different directions through adaptive horizontal pooling and vertical pooling operations to obtain features of the rectangular region, resulting in two feature maps of different sizes, denoted as . and ;Will and Adding them together yields the feature map after remodeling the rectangular region, denoted as . Then, a weight map is generated using the Sigmoid activation function, denoted as . ;Will and Element-wise multiplication yields the output of the rectangular self-calibration module, denoted as... ;against and Use and obtain The same method is used to obtain the output of the rectangular self-calibration module, denoted as... and ; By adaptive average pooling , , Feature maps adjusted to the same size are denoted as , , ;right , , The result of feature concatenation and enhancement is denoted as: Then Perform channel-dimensional segmentation, and denote the output enhanced features as follows: , , ;Will and The data is fed into the multi-scale feature fusion module, and the fused features are denoted as follows: ;Will and The fusion is performed in the dynamic interpolation fusion module, and the fused features are denoted as follows. ;Will and After fusion by the multi-scale feature fusion module, the fused features are denoted as follows: ;Will and The fusion is performed in the dynamic interpolation fusion module, and the fused features are denoted as follows. ;Will and The feature output is calculated by the multi-scale feature fusion module, and the output feature is denoted as . ; Will , , The data is fed into the detection head, and the location information of the small bird target is calculated.

3. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 2, characterized in that: The The feature map is obtained by feeding it into the C2f module, which is an improvement based on the multi-scale edge enhancement module, and denoted as . ,include: First, the size is gradually reduced by using the initial convolutional layers and spatial channel downsampling of the YOLOv10 backbone network, while extracting basic features. The resulting feature map is denoted as... Then on Adaptive average pooling is performed, the number of channels is reduced through convolution operations, and then spatial details are extracted through standard convolution operations to obtain multi-scale features, denoted as . ; Average pooling is used for multi-scale features Smoothing is performed to generate low-frequency features, denoted as... ;right and Performing difference operations preserves high-frequency information in rapidly changing regions of the image, yielding high-frequency features, denoted as... ; Channel information is fused through convolution operations, and a high-frequency feature weight map is generated using the Sigmoid activation function, which is then combined with... Superposition yields the output features enhanced with high-frequency information, denoted as... ; Feature maps processed by the edge enhancement module at different scales Bilinear interpolation is performed to upsample to the input size to maintain spatial alignment, resulting in the output of each branch, denoted as... For multi-scale features The local features are obtained through convolution operations and denoted as follows: ; Will The enhanced features are concatenated with the outputs of each branch along the channel dimension to obtain the enhanced features, denoted as . ; Next to Perform average pooling and max pooling operations to obtain the channel-compressed feature map, denoted as . ; Depth-separable convolution pairs Deep feature extraction is performed, and residual connections are used to avoid excessive suppression of original features by spatial attention. The resulting features are denoted as follows: ; Will The low-frequency components obtained through global average pooling are denoted as follows: and with Perform a difference operation to obtain the high-frequency component, denoted as . Next, and Residual modulation is performed to obtain an enhanced feature map, denoted as... ; Will The number of channels is adjusted through convolution operations to obtain the final output features. .

4. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 2, characterized in that: The The third layer of features is obtained through the large-kernel spatial attention module, specifically: Extracting from convolutional branches with three different receptive fields Feature maps with different receptive field ranges, denoted as , , ; right , , The correlation between channels was modeled using convolution operations to obtain... ; Connect along the channel dimension using matrix concatenation operations. The fused feature map is obtained and denoted as , ; Again Perform average pooling and max pooling operations to obtain spatial statistics, denoted as... ; Split along channel dimension We obtain three attention weights, denoted as... ; The generated attention weights are multiplied element-wise with the corresponding branch features to obtain the fused features, which are then restored through convolution operations and summed. Multiplying each element together yields the third layer of features.

5. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 2, characterized in that: The and The data is fed into the multi-scale feature fusion module, including: for First, features are extracted using convolutional layers. Then, the output range of the features is adjusted using the sigmoid activation function. The adjusted feature map is denoted as... ;Will and Perform element-wise multiplication to obtain ; The and The fusion is performed in the dynamic interpolation fusion module, including: firstly, ... Perform interpolation upsampling and adjust its size to match... The same, after further convolution operations and Integration, to obtain .

6. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 1, characterized in that: The step of feeding the location information of small bird targets in both the binocular and left-eye images into the backbone feature extraction network to extract multi-scale features includes: For the left eye image in a binocular image, the generated multi-scale feature map is denoted as... ; Combine the location information of small bird targets in the left eye image with The resulting features are denoted as follows: ; Will The decoder outputs an enhanced multi-scale feature map, denoted as... ; Used Using the same method, for the right eye image in a binocular image, an enhanced multi-scale feature map is obtained, denoted as... .

7. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 6, characterized in that: The construction of a multi-level cost body from the extracted multi-scale features includes: Select and Construct an initial cost body, use the dot product of feature vectors as a similarity measure, and calculate... Each pixel and The matching cost for all pixels in the corresponding row is used to construct a 4D correlation cost volume; the 4D cost volume is then converted into a 3D correlation cost volume by calculating the feature dot product under the same y coordinate only. Then, three one-dimensional average pooling operations are performed on the 3D related cost volume to obtain a four-layer multi-level cost volume as the final multi-level cost volume.

8. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 7, characterized in that: The disparity iterative update calculation based on the multi-frequency channel attention mechanism for the constructed multi-level cost volume includes: Will Contextual features are generated using the ReLU activation function, denoted as […]. Set initial parallax Initiate the parallax iteration process; for First, preprocessing is performed using convolutional layers to obtain... High- and low-frequency separation is achieved by converting contextual features to the frequency domain through two-dimensional discrete cosine transform. Next, the Top-K algorithm is used to select the top K key components with the greatest impact on matching accuracy from all frequency domain components based on a predefined frequency index. Redundant low-frequency noise and invalid high-frequency interference are eliminated, balancing feature representation and computational efficiency. Statistical information is extracted from the selected K frequency feature maps using global average pooling, global max pooling, and global min pooling, respectively, resulting in three statistical vectors, denoted as . The three statistical vectors are first passed through a dimension-reduction fully connected layer, then processed by the ReLU activation function, then through a dimension-upgrading fully connected layer to restore the number of channels, and finally the output is mapped to the [0,1] interval by the Sigmoid activation function to generate frequency attention weights, denoted as . ; Will The hidden information of the gated recurrent unit at the corresponding level is multiplied element by element to achieve adaptive enhancement of information of different frequencies, and the hidden information of the gated recurrent unit after correction by the multi-frequency channel attention mechanism is obtained.

9. The method for detecting and locating small bird targets under location awareness and frequency enhancement according to claim 1, characterized in that: The method of supervising and optimizing disparity results through a loss function to achieve spatial localization of small bird targets includes: The L1 loss function is used to supervise the disparity iteration process, and the network parameters are optimized through backpropagation to reduce disparity error. The optimized disparity results are upsampled to restore the original image resolution, resulting in the final disparity map; The intrinsic and extrinsic parameters of the binocular camera are calculated, the depth information of small bird targets is calculated based on the principle of triangulation, and then the three-dimensional spatial coordinates of the target are obtained through coordinate transformation to achieve precise positioning.

10. A bird small target detection and localization system based on position awareness and frequency enhancement, characterized in that, include: The data acquisition module is used to acquire a dataset of small bird target images in substations, wherein the dataset of small bird target images in substations is a binocular image dataset; The bird target location information acquisition module is used to input the left eye image from the binocular images into the neural network to obtain the location information of bird targets in the left eye image; The multi-scale feature extraction module is used to input the location information of small bird targets in the binocular image and the left eye image into the backbone feature extraction network to extract multi-scale features. The multi-level cost body construction module is used to construct a multi-level cost body from the extracted multi-scale features. The disparity iterative update module is used to perform disparity iterative update calculations on the constructed multi-level cost volume based on a multi-frequency channel attention mechanism. The bird small target localization module is used to achieve spatial localization of bird small targets by supervising and optimizing the disparity results through a loss function.

Images