Bird fine-grained detection algorithm based on snakelike deformable convolution and cavity convolution

Abstract

The invention discloses a bird fine-grained detection algorithm based on snakelike deformable convolution and cavity convolution, and relates to bird recognition. The technical problems that according to an existing bird fine-grained recognition technology, fine-grained feature extraction and positioning coupling is insufficient, target form learning is difficult, the modeling capacity of local deformation is limited, fine-grained sensing is limited, and deployment resource consumption is huge are solved. The algorithm comprises the following steps: S11, extracting a predicted dynamic offset according to bird image information; s12, performing dynamic deformation according to the dynamic offset to generate a target sampling coordinate; s13, obtaining bird characteristics after dynamic deformation through the target sampling coordinates; s14, processing according to the dynamically deformed bird characteristics to obtain enhanced bird characteristics; s15, fusing the enhanced bird features and the basic bird features to generate bird fusion features; and S16, outputting fine-grained features of the target birds according to the fused features of the birds. According to the method, the parameter quantity and the calculation complexity of the model are reduced, the lightweight requirement is met, and local deformation and fine features are effectively captured.

Description

Technical Field

[0001] The present invention relates to bird recognition, and more particularly, to a fine-grained bird detection algorithm based on snake-like deformable convolution and dilated convolution. Background Art

[0002] Bird species identification, as a key technical tool for biodiversity monitoring and ecological conservation, has garnered widespread attention in recent years. Research has shown a significant correlation between bird population dynamics and habitat changes, and real-time monitoring of bird colony structure can effectively assess ecosystem health. Traditional bird identification relies primarily on manual observation and acoustic analysis, but this suffers from limitations such as low efficiency and subjectivity in large-scale monitoring scenarios. With the advancement of computer vision technology, bird image classification based on deep learning has gradually become a mainstream solution.

[0003] Current deep learning-based fine-grained bird recognition technology is mainly optimized around the target detection framework and can be divided into three major technical routes:

[0004] CNN single-stage target detection method: Representative works such as YOLOv5 and RetinaNet directly predict bird locations and species categories through end-to-end regression.

[0005] CNN multi-stage object detection methods: A two-stage framework, represented by Faster R-CNN, generates candidate boxes through a region proposal network (RPN) and then performs fine-grained classification. Transformer object detection methods: DETR first introduced the Transformer encoder-decoder architecture to the detection task, capturing global feature dependencies through a self-attention mechanism.

[0006] Defects and shortcomings of existing technology:

[0007] 1. Insufficient coupling between fine-grained feature extraction and positioning: Existing single-stage methods (such as YOLOv5 and SSD) do not explicitly decouple fine-grained semantic information from spatial positioning features during the feature extraction process, resulting in key discriminative features such as feather texture and beak morphology being easily interfered with by background noise.

[0008] 2. Difficulty learning object morphology and limited modeling capabilities for local deformation: The RPN network used by Faster R-CNN generally uses fixed-scale anchor boxes to generate candidate regions. However, for deformed or irregular objects, these fixed-shape anchor boxes often have difficulty effectively matching the object shape, resulting in missed detections. Because traditional R-CNN relies on RoIPooling or RoIAlign, its ability to model local deformation in object detection is limited.

[0009] 3. The Transformer architecture has limited fine-grained perception and consumes huge deployment resources: Traditional Transformers rely on global attention (Self-Attention). Although they can effectively learn long-distance dependencies, they are weak in capturing local deformation features, resulting in unstable performance of the detector for small and edge targets. The computational complexity is too high and difficult to deploy on lightweight devices: The classic DETR needs to calculate global attention between all pixels, resulting in a computational complexity of O(N 2 ), under high-resolution input, the inference time is long and it is difficult to apply to real-time scenarios. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to address the shortcomings of the existing technology and provide a bird fine-grained detection algorithm based on snake-shaped deformable convolution and void convolution to solve the technical problems of the existing bird fine-grained recognition technology, such as insufficient coupling between fine-grained feature extraction and positioning, difficulty in learning target morphology, limited modeling ability of local deformation, limited fine-grained perception and huge deployment resource consumption.

[0011] The present invention discloses a bird fine-grained detection algorithm using serpentine deformable convolution and dilated convolution, which includes obtaining bird image information, triggering a first convolution strategy based on the bird image information to output bird fusion features, and triggering a target convolution strategy based on the bird fusion features to output target bird fine-grained features.

[0012] As a further improvement, the first convolution strategy includes:

[0013] S11: extracting basic bird features from the bird image information, and predicting a dynamic offset based on the basic bird features;

[0014] S12: constructing a basic sampling network, and dynamically deforming the basic sampling coordinates of the basic sampling network according to the dynamic offset to generate target sampling coordinates;

[0015] S13: performing bilinear difference sampling on the basic bird features using the target sampling coordinates to obtain dynamically deformed bird features;

[0016] S14: performing subsequent convolution, normalization, and activation processing on the dynamically deformed bird features to obtain enhanced bird features;

[0017] S15: Fusing the enhanced bird feature with the basic bird feature to generate a bird fusion feature.

[0018] Furthermore, in S11, the method for predicting the dynamic offset is to set a convolution layer, predict a preliminary offset based on the convolution layer and basic bird features, and map the preliminary offset value to a dynamic offset through a Batch Normalization data normalization function and a tanh activation function.

[0019] Furthermore, in S12, the method for generating the target sampling coordinates is to set the original center coordinates and the convolution kernel template to form a basic sampling network, obtain the basic sampling coordinates of the basic sampling network, and apply dynamic offsets to the horizontal coordinates and vertical coordinates of the basic sampling coordinates according to the direction of the dynamic deformation to generate the target sampling coordinates.

[0020] Furthermore, when the direction of the dynamic deformation is vertical, the expression for applying the dynamic offset to the vertical coordinate is:

[0021] y base =y center +0,x base =x center +x grid ;

[0022]

[0023] x new =x base ;

[0024] Among them, (x base ,y base ) represents the basic sampling coordinates, (x center ,y center ) represents the original center coordinates, the center position of the convolution kernel When the dynamic deformation direction is set to vertical, the x component of the convolution kernel template is in the interval Evenly distributed within, Indicates the dynamic offset of the vertical coordinate, pass Calculated, extend_scope is used for scaling Determine the action offset, (x new ,y new ) represents the target sampling coordinates.

[0025] Furthermore, when the direction of the dynamic deformation is horizontal, a dynamic offset is applied to the horizontal coordinate. The expression for applying the dynamic offset to the horizontal coordinate is:

[0026] y base =y center +0,x base =xcenter +x grid ;

[0027] y new =y base ;

[0028] Among them, (x base ,y base ) represents the basic sampling coordinates, (x center ,y center ) represents the original center coordinates, the center position of the convolution kernel When the dynamic deformation direction is set to horizontal, the y component of the convolution kernel template is in the interval Evenly distributed within, Indicates the dynamic offset of the horizontal coordinate, pass Calculated, extend_scope is used to scale by The dynamic offset obtained, (x new ,y new ) represents the target sampling coordinates.

[0029] Furthermore, in S13, the expression of the bird characteristics after dynamic deformation is obtained as follows:

[0030] f(y new x new )≈f(y0,x0)·(y1-y)(x1-x)+f(y0,x1)·(y1-y)(x-X0)

[0031] +f(y1,x0)·(y-y0)(x1-x)+f(y1,x1)·(y-y0)(x-x0);

[0032] in, y1=y0+1, x1=x0+1,f(y new ,x new ) represents the characteristics of the bird after dynamic deformation.

[0033] Furthermore, in S15, the expression for generating the bird fusion feature is:

[0034] Output1=concat(conv_0(x), conv_x(x), conv_y(x))

[0035] Among them, Output1 is the bird fusion feature, conv_0(x) is the basic bird feature, conv_x(x) is the feature quantity of the enhanced bird feature in the horizontal direction, and conv_y(x) is the feature quantity of the enhanced bird feature in the vertical direction.

[0036] Furthermore, the target convolution strategy includes:

[0037] S161: generating main features of the main convolution branch according to the bird fusion features;

[0038] S162: Generate cheap features of cheap convolution branches according to the bird fusion features;

[0039] S163: The main features and the cheap features are spliced together to output fine-grained features of the target bird.

[0040] Furthermore, in S163, the expression of the target bird fine-grained feature is spliced and output as follows:

[0041] Output2=concat(y,z),Output2∈R B×c2×H′×W ';

[0042] Among them, Output2 represents the fine-grained features of the target bird, y represents the main features, z represents the cheap features, and B×c2×H′×W′ represents the size of the fine-grained features of the target bird in the channel dimension.

[0043] Beneficial effects

[0044] The advantages of the present invention are:

[0045] The present invention obtains bird image information, triggers the first convolution strategy according to the bird image information to output bird fusion features, and triggers the target convolution strategy according to the bird fusion features to output target bird fine-grained features. Through deep grouped convolution and dynamic sampling, the number of parameters and computational complexity of the model are reduced, meeting lightweight requirements, effectively capturing local deformation and subtle features, improving classification and detection accuracy in complex scenarios, ensuring feature richness while significantly reducing redundant calculations, and having high stability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the overall neural network structure of the first convolution strategy and feature fusion of the present invention;

[0047] Figure 2 Schematic diagram of the standard convolution branch neural network structure in the first convolution strategy of the present invention;

[0048] Figure 3This is a diagram of the dynamic convolution branch X neural network structure in the first convolution strategy of the present invention;

[0049] Figure 4 This is a neural network structure diagram of the dynamic convolution branch Y in the first convolution strategy of the present invention;

[0050] Figure 5 This is a neural network structure diagram of the target convolution strategy of the present invention;

[0051] Figure 6 This is a comparison chart of the training results of the present invention and other neural networks. DETAILED DESCRIPTION

[0052] The present invention will be further described below in conjunction with the embodiments, but this does not constitute any limitation to the present invention. Any limited number of modifications made by anyone within the scope of the claims of the present invention are still within the scope of the claims of the present invention.

[0053] See Figures 1-6 The present invention provides a bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution, which includes obtaining bird image information, triggering a first convolution strategy according to the bird image information to output bird fusion features, and triggering a target convolution strategy according to the bird fusion features to output target bird fine-grained features.

[0054] The first object of the present invention is to provide a dynamic convolution method called DualAxis-Deform-Conv (dual-axis-deformation-convolution), which uses the DualAxis-Deform-Conv (dual-axis-deformation-convolution) dynamic convolution method as the first convolution strategy, so that the algorithm can obtain stronger feature expression capabilities when processing targets with deformation or irregular boundaries, while also achieving efficient capture of local deformation information in different directions.

[0055] Detailed process analysis of DualAxis-Deform-Conv (dual-axis-deformation-convolution) design and feature fusion:

[0056] 1. Obtain bird image information, extract basic bird features from the bird image information, and predict dynamic offsets based on these basic bird features. Dynamic offset prediction (using 3x3 convolution + BN + tanh) limits the offset value to [-1, 1].

[0057] For the input feature f∈R B×C×W×H ,

[0058] ①First, a 3x3 convolutional layer is used to predict the initial offset: Δ∈R B×(2K)×W×H , where K is the size of the convolution kernel.

[0059] ②Then the output is divided into two parts in the channel dimension: Δ=[Δ y , Δ x ], the size of each part is (B×K×W×H)

[0060] ③ Then, the initial offset value is mapped to [-1, 1] through Batch Normalization and tanh activation function to form a dynamic offset:

[0061] Δ←tanh(BN(Δ)).

[0062] 2. Build a basic sampling network, such as Figure 1 As shown, the base sampling coordinates of the base network are dynamically deformed according to the dynamic offset to generate the target sampling coordinates. According to the preset convolution kernel template (horizontally or vertically), the dynamic offset is accumulated and the sampling coordinates are adjusted.

[0063] For each pixel location (p, q), we first construct a basic sampling network based on the original center coordinates (x center ,y center ) and the preset convolution kernel template:

[0064] ①Convolution kernel template

[0065] Depending on the direction of dynamic deformation:

[0066] (1) When morph = 1 (vertical direction):

[0067] The y component in the template is fixed to 0, and the x component is in the range The internal uniform distribution is:

[0068] y grid＝0 ,

[0069] (2) When morph = 0 (horizontal direction):

[0070] The x component is fixed to 0 in the template, and the y component is The internal uniform distribution is:

[0071] x grid =0,

[0072] ② Dynamic offset (snake accumulation)

[0073] Dynamically adjust the base grid using the dynamic offset to generate the target sampling coordinates, taking the vertical direction (morph=1) as an example:

[0074] (1) For each convolution kernel point, its basic sampling coordinates are:

[0075] ybase =y center +0,x base =x center +x grid

[0076] (2) Apply a dynamic offset to the y coordinate, such as Figure 4 As shown, the center position of the convolution kernel As a benchmark, it is stipulated that:

[0077]

[0078] At the same time, the sampling coordinates are updated in the reverse direction (to ensure the center position is fixed).

[0079]

[0080] Indicates the dynamic offset of the vertical coordinate, pass Calculated, extend_scope is used for scaling The determined action offset is used to adjust the basic sampling coordinates in the vertical direction to generate the target coordinates in the vertical direction, (x new ,y new ) represents the target sampling coordinates.

[0081] While the x coordinate remains:

[0082] x new =x base .

[0083] like Figure 3 As shown, for the horizontal direction (morph=0), a similar process is used to dynamically offset the x coordinate, and the formula is:

[0084]

[0085] The y coordinate remains y new =y base .

[0086] in, Indicates the dynamic offset of the horizontal coordinate, pass Calculated, extend_scope is used to scale by The obtained dynamic offset is used to adjust the basic sampling coordinates in the horizontal direction, and then generate the target coordinates in the horizontal direction, (x new ,y new ) represents the target sampling coordinates.

[0087] 3. Bilinear interpolation sampling is used to obtain enhanced bird features at the bird feature sampling points after dynamic deformation.

[0088] Using the target sampling coordinates (x new ,y new ) Perform bilinear interpolation sampling on the input feature f:

[0089] in y1=y0+1, x1=x0+1,f(y new ,x new ) represents the characteristics of the bird after dynamic deformation. This sampling method ensures the continuity and smoothness of the feature map, making the characteristics after dynamic deformation more adaptable.

[0090] 4. Perform subsequent convolution, normalization, and activation processing on the dynamically deformed bird features to obtain enhanced bird features: After obtaining the dynamically deformed bird feature map:

[0091] (1) Call the convolution operation with a kernel size of (K, 1) along the horizontal branch (morph = 0),

[0092] (2) Call the convolution operation with a convolution kernel size of (1, K) along the vertical branch (morph=1).

[0093] Next, the output will go through Group Normalization (GN) and activation function ReLU to enhance nonlinear expression and numerical stability.

[0094] 5. Fusion and final output

[0095] In the forward propagation of the DualAxis-Deform-Conv convolution method:

[0096] (1) Standard convolution branch: directly extract basic features.

[0097] (2) Dynamic snake convolution branch: The above dynamic offset and sampling mechanism is used to generate deformed features in the horizontal (conv_x) and vertical (conv_y) directions respectively.

[0098] (3) Fusion: Fusion output, splicing three-way features to obtain a richer representation.

[0099] The general steps are as follows:

[0100] Three branches are defined in the constructor of DualAxisDeformConv:

[0101] ① Standard convolution branch (conv_0): uses conventional convolution (e.g. Conv(inc, ouc, k)) to extract basic features without involving dynamic offset.

[0102] ② Dynamic convolution branch (two branches): conv_x and conv_y. Their internal mechanisms were originally implemented in DSConv, but are now directly embedded in DySnakeConv:

[0103] (1) conv_x: Dynamic sampling and convolution in the horizontal direction (along the x-axis), passing in the parameter morph = 0;

[0104] (2) conv_y: Dynamic sampling and convolution in the vertical direction (along the y-axis), passing in the parameter morph = 1.

[0105] During forward propagation, the outputs of these three branches are concatenated in the channel dimension, and the formula is:

[0106] Output1=concat(conv_0(x), conv_x(x), conv_y(x));

[0107] This design preserves the information of standard convolution while integrating adaptive dynamic features in two directions, making the network more adaptable to local deformation and edge irregularities. The fusion result contains both the global information captured by standard convolution and the sensitivity to local geometric deformation brought by dynamic sampling.

[0108] The target convolution strategy is triggered based on the bird fusion features to output the fine-grained features of the target bird. The present invention provides a Split-Efficient-Conv convolution method, which uses the Split-Efficient-Conv convolution method as the target convolution strategy. While ensuring computational efficiency, it can obtain feature representations with sufficient expressive power at a low computational cost and fully utilize redundant information, thereby achieving efficient feature learning and optimizing computing resources.

[0109] Detailed process analysis of the convolution method of Split-Efficient-Conv: 1. Main convolution operation, generating preliminary features through smaller convolution kernels; 2. Cheap convolution operation, using larger convolution kernels and depthwise convolution (grouped convolution) to generate additional features; 3. Activation function performs nonlinear transformation on the features after convolution operation; 4. Concatenate the features of the main convolution branch and the cheap convolution branch to obtain a comprehensive feature representation; 5. The final output has a feature representation with small computational complexity and powerful functions, low time complexity and the ability to enrich the feature space.

[0110] Its general module structure is as follows:

[0111] The main features of the main convolution branch are generated according to the bird fusion features, the cheap features of the cheap convolution branch are generated according to the bird fusion features, and the main features and the cheap features are spliced to output the fine-grained features of the target bird.

[0112] The Split-Efficient-Conv module accepts the number of input channels c1, the number of output channels c2, as well as the convolution kernel size k, stride s, number of groups g, and activation flag act. The module mainly contains two convolution operations:

[0113] 1. Main convolution branch cV1

[0114] Used to generate the main features, the number of output channels is:

[0115] The parameters used when calling are: input channel c1, output channel c′, convolution kernel size k, stride s, number of groups g, and whether to activate act.

[0116] Mathematically, it can be described as: y = σ(Conv1(x; W1)), y∈R B×C′×H′×W′ ;

[0117] Where σ represents the activation function (if act=True) and σ is the convolution kernel parameter.

[0118] Calculate the main features

[0119] ① Input: Input feature x∈R B×c1×H×W ;

[0120] ②Main convolution: Use cv1 to perform convolution operation, y=Conv1(x).

[0121] Note: The parameters used here include the convolution kernel size k and stride s, the number of groups is g, and the number of output channels is c′=c2 / 2.

[0122] ③Activation: If act is true, an activation function will be added after the convolution, so that the output has nonlinear expression capabilities.

[0123] 2. Cheap convolution branch cv2

[0124] It is used to generate more ("ghost") features from the main features. Its design uses a larger convolution kernel (5x5 here) and the idea of deep convolution (the number of groups is set to the number of input channels, i.e. c_') to ensure the lowest computation time.

[0125] Mathematically described as: z = σ(Conv2(y; W2)), z∈RB×C′×H′×W′ ;

[0126] Here, the parameter W2 of Conv2 uses a convolution kernel size of 5, a stride of 1, and a grouping number of c′, thus achieving an effect similar to depthwise convolution.

[0127] ① Cheap convolution: After obtaining the main feature y, y is processed using CV2, z = Conv2(y). CV2 uses a larger convolution kernel (5x5) and a stride of 1, along with a grouping of 2 (deep convolution), to minimize computational overhead. This computation is primarily used to generate "ghost" features, or cheap features, that are redundant with y but provide supplementary information.

[0128] ②Activation: Similarly, according to the act parameter, the activation function is added after the convolution.

[0129] Finally, the output of the module is the concatenation of the features of the main convolution branch and the cheap convolution branch along the channel dimension:

[0130] Output2=concat(y,z),Output∈R B×(c′+c′)×H′×W′ =R B×c2×H′×W′ .

[0131] Finally, the main feature y and the cheap feature z are concatenated in the channel dimension, Output2 = concat(y, z). After concatenation, the number of output channels is c′+c′=c2, achieving the expected output dimension.

[0132] The above is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several modifications and improvements can be made without departing from the structure of the present invention. These modifications and improvements will not affect the effect of the implementation of the present invention and the practicality of the patent.

Claims

1. A fine-grained bird detection algorithm based on snake-shaped deformable convolution and dilated convolution, characterized by: The algorithm includes, S11: Acquire bird image information, extract basic bird features from the bird image information, and predict a dynamic offset based on the basic bird features; S12: constructing a basic sampling network, and dynamically deforming the basic sampling coordinates of the basic sampling network according to the dynamic offset to generate target sampling coordinates; S13: sampling basic bird features using the target sampling coordinates to obtain dynamically deformed bird features; S14: processing the dynamically deformed bird features to obtain enhanced bird features; S15: fusing the enhanced bird feature with the basic bird feature to generate a bird fusion feature; S16: triggering a target convolution strategy according to the bird fusion features to output target bird fine-grained features.

2. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 1 is characterized in that: In S12, the method for generating the target sampling coordinates is to set the original center coordinates and the convolution kernel template to form a basic sampling network, obtain the basic sampling coordinates of the basic sampling network, and apply dynamic offsets to the horizontal coordinates and vertical coordinates of the basic sampling coordinates according to the direction of the dynamic deformation to generate the target sampling coordinates.

3. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 2 is characterized in that: When the direction of the dynamic deformation is vertical, a dynamic offset is applied to the vertical coordinate. The expression for applying the dynamic offset to the vertical coordinate is: and base =and center +0,x base =x center +x grid ; x new =x base ; Among them, (x base ,y base ) represents the basic sampling coordinates, (x center ,y center ) represents the original center coordinates, the center position of the convolution kernel When the dynamic deformation direction is set to vertical, the x component of the convolution kernel template is in the interval Evenly distributed within, Indicates the dynamic offset of the vertical coordinate, pass Calculated, extend_scope is used for scaling Determine the action offset, (x new ,y new ) represents the target sampling coordinates.

4. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 2 is characterized in that: When the direction of the dynamic deformation is horizontal, a dynamic offset is applied to the horizontal coordinate. The expression for applying the dynamic offset to the horizontal coordinate is: and base =and center +0,x base =x center +x grid ; and new =and base ; Among them, (x base ,y base ) represents the basic sampling coordinates, (x center ,y center ) represents the original center coordinates, the center position of the convolution kernel When the dynamic deformation direction is set to horizontal, the y component of the convolution kernel template is in the interval Evenly distributed within, Indicates the dynamic offset of the horizontal coordinate, pass Calculated, extend_scope is used to scale by The dynamic offset obtained, (x new ,y new ) represents the target sampling coordinates.

5. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 1 is characterized in that: In S13, bilinear interpolation sampling is performed on the basic bird features to obtain the dynamically deformed bird features. The expression of the dynamically deformed bird features is: f(y new x new )≈f(y0,x0)·(y1-y)(x1-x)+f(y0,x1)·(y1-y)(x-x0)+f(y1,x0)·(y-y0)(x1-x)+f(y1,x1)·(y-y0)(x-x0); in, y1=y0+1, x1=x0+1,f(y new , x new ) represents the characteristics of the bird after dynamic deformation.

6. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 1 is characterized in that: In S15, the expression for generating the bird fusion feature is: Output1=concat(conv_0(x), conv_x(x), conv_y(x)); Among them, Output1 is the bird fusion feature, conv_0(x) is the basic bird feature, conv_x(x) is the feature quantity of the enhanced bird feature in the horizontal direction, and conv_y(x) is the feature quantity of the enhanced bird feature in the vertical direction.

7. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 1 is characterized in that: In S11, the method for predicting the dynamic offset is to set a convolution layer, predict a preliminary offset based on the convolution layer and basic bird features, and map the preliminary offset value to a dynamic offset through a Batch Normalization data normalization function and a tanh activation function.

8. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 1 is characterized in that: In S14, subsequent convolution, normalization and activation processing are performed on the dynamically deformed bird features to obtain enhanced bird features.

9. The bird fine-grained detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 1 is characterized in that: In S16, the target convolution strategy includes: S161: generating main features of the main convolution branch according to the bird fusion features; S162: Generate cheap features of cheap convolution branches according to the bird fusion features; S163: The main features and the cheap features are spliced together to output fine-grained features of the target bird.

10. The fine-grained bird detection algorithm based on snake-shaped deformable convolution and dilated convolution according to claim 9 is characterized in that: In S163, the expression of the target bird fine-grained feature is spliced and output as follows: Output2=concat(y,z),Output2∈R B×c2×H′×W′ ; Among them, Output2 represents the fine-grained features of the target bird, y represents the main features, z represents the cheap features, and B×c2×H′×W′ represents the size of the fine-grained features of the target bird in the channel dimension.

Images