YOLOV5 bird detection method based on dynamic snakelike convolution
By introducing a dynamic serpentine convolution module in YOLOV5, the diversification and dynamic changes in bird recognition are solved, the accuracy and stability of bird recognition are improved, and the ability to classify birds is enhanced.
Patent Information
- Application Number
- CN202510461611.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-14
- Publication Date
- 2025-08-08
AI Technical Summary
The prior art has the problem of high recognition difficulty in bird recognition, including factors such as diverse bird appearance, changing posture, rapid movement and distance changes, which makes it difficult to guarantee the accuracy and real-time recognition.
The dynamic serpentine convolution module is used to replace the C3 convolution module in YOLOV5, and a dynamic serpentine convolution module is built to improve the body contour capture ability and recognition accuracy of birds through training models.
It improves the accuracy and stability of bird recognition, can effectively capture the body contour and posture information of birds, adapt to changes in spatial structure in the image, and enhances the ability to classify birds.
Smart Images
Figure CN120452013A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of flying bird image recognition, and in particular relates to a YOLOV5 flying bird detection method based on dynamic snake convolution. Background Art
[0002] With the advancement of modern society, the aviation sector has experienced rapid development, with increasing emphasis on using computer technology to enhance flight safety. Bird strikes have become a significant threat to aviation safety. However, by improving the accuracy of bird target recognition, timely measures can be taken to repel birds, effectively reducing the harmful effects of bird strikes.
[0003] Therefore, modern technologies are being applied to the field of bird target recognition, providing more accurate and reliable bird identification capabilities for aviation. By using technologies such as computer vision and deep learning, it is possible to automatically identify and monitor birds in flight. However, due to the inherent characteristics of birds and interference from the external environment, the following identification difficulties exist:
[0004] 1. Birds are very diverse in appearance and posture. Different species of birds vary greatly in size, shape, color, etc. In addition, birds may adopt different postures during flight, such as flapping wings, gliding, circling, etc., which increases the difficulty of identification.
[0005] 2. Birds often move quickly and have varying trajectories while in flight, making real-time processing of video or continuous frame images a challenge. Furthermore, birds may fly at varying depths, requiring algorithms to effectively handle changes in distance.
[0006] 3. The key to bird recognition is extracting and selecting effective visual features for classification. Different bird types are primarily identified based on features such as color, texture, and shape. Therefore, designing accurate feature extraction methods is a key challenge in bird recognition.
[0007] Therefore, based on the above difficulties in the prior art, the present invention discloses a YOLOV5 flying bird detection method based on dynamic snake convolution. Summary of the Invention
[0008] The present invention discloses a YOLOV5 flying bird detection method based on dynamic snake convolution. By introducing dynamic snake convolution, the model's ability to capture the body contours of flying birds is improved, thereby improving the model's ability to classify flying birds.
[0009] The present invention is achieved through the following technical solutions:
[0010] A YOLOv5 bird detection method based on dynamic snake convolution is proposed. The dynamic snake convolution module is used to replace the C3 convolution module in YOLOv5. Then, the collected bird images are input for training to obtain an optimal bird recognition model. Bird recognition is achieved through the optimal bird recognition model. The method includes the following steps:
[0011] Step 1: Construct training and validation sets for bird recognition.
[0012] Step 2: Build a dynamic snake convolution module. Use the dynamic snake convolution module to replace the C3 convolution module in YOLOv5 to obtain a training model. The dynamic snake convolution module includes at least two parallel convolution routes, each of which has at least one convolution layer, and at least one convolution route has a dynamic snake residual module. The convolution results of all convolution routes are aggregated and then input into the convolution layer again.
[0013] Step 3: Input the collected bird images into the training model in batches and iterate to obtain the optimal bird recognition model;
[0014] Step 4: Implement flying bird recognition through the optimal flying bird recognition model.
[0015] In order to better implement the present invention, further, the dynamic snake-shaped residual module includes a base convolution layer, a dynamic snake-shaped convolution layer, and an add module connected in sequence, and the add module is used to add the input of the base convolution layer and the output of the dynamic snake-shaped convolution layer as the output of the dynamic snake-shaped convolution module.
[0016] In order to better implement the present invention, further, the dynamic serpentine convolution module includes a first convolution layer, a second convolution layer, a dynamic serpentine residual module, an aggregation function module, and a third convolution layer. The first convolution layer is connected to the dynamic serpentine residual module, and the dynamic serpentine residual module is used to aggregate the output of the dynamic serpentine residual module with the output of the second convolution layer and input them into the third convolution layer.
[0017] In order to better implement the present invention, further, the input of the first convolutional layer and the second convolutional layer is the same, denoted as λ; the λ obtains a first output after passing through the first convolutional layer and the dynamic snake-shaped residual module, and the λ obtains a second output after passing through the second convolutional layer; the aggregation function module concats the first output and the second output into a third output, and the third output is input to the third convolutional layer for processing as the final output of the dynamic snake-shaped convolution module.
[0018] In order to better implement the present invention, further, the convolution kernel size of the first convolution layer is 1×1, the stride is 1, and the padding is 0.
[0019] In order to better implement the present invention, further, the convolution kernel size of the second convolution layer is 1×1, the stride is 1, and the padding is 0.
[0020] In order to better implement the present invention, further, the convolution kernel size of the third convolutional layer is 1×1, the stride is 1, and the padding is 0.
[0021] In order to better implement the present invention, further, the convolution kernel size of the basic convolution layer is 1×1, the stride is 1, and the padding is 0.
[0022] In order to better implement the present invention, further, the step 3 specifically includes:
[0023] Step 3.1: Input the flying bird images in the training set into the training model in batches to obtain a training model set;
[0024] Step 3.2: Input the flying bird images in the validation set into the training model set and calculate the MAP index of the validation set to obtain the MAP index set;
[0025] Step 3.3: Extract the maximum index value from the MAP index set, and extract the training model corresponding to the maximum index value from the training model set as the optimal flying bird recognition model.
[0026] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0027] 1. Birds often have long, narrow bodies. The dynamic snake convolution module in this invention can help identify this long, narrow feature. By operating along a serpentine path during the convolution process, the dynamic snake convolution module can better capture the bird's body contours and improve recognition accuracy.
[0028] 2. The dynamic snake convolution module in this invention can flexibly adjust the snake path during the convolution process, selecting the appropriate path based on the characteristics of different areas in the image. This enables the dynamic snake convolution module to better perceive the spatial structure in the image for the bird recognition task, improving the ability to recognize information such as the bird's posture and flight direction;
[0029] 3. The dynamic snake convolution module in the present invention has a certain robustness against some distortions and scale changes in the image. Flying birds may have some deformation or scale scaling in actual scenes. The dynamic snake convolution module can better cope with these changes and improve the stability of recognition. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 Schematic diagram of the process of bird detection method;
[0031] Figure 2Schematic diagram of the dynamic snake convolution module;
[0032] Figure 3 Schematic diagram of the dynamic snake residual module. DETAILED DESCRIPTION
[0033] Example 1:
[0034] A YOLOV5 bird detection method based on dynamic snake convolution in this embodiment is as follows: Figure 1 As shown in the figure, the dynamic snake convolution module is used to replace the C3 convolution module in YOLOv5, and then the collected bird images are input for training to obtain the optimal bird recognition model. The optimal bird recognition model is used to implement bird recognition, including the following steps:
[0035] Step 1: Construct a training set and a validation set for bird recognition. Collect N types of bird images, select K bird images as the training set, and the remaining bird images as the validation set, where K < N.
[0036] Step 2: Build a dynamic snake convolution module, denoted as: C3_DySnake. Use C3_DySnake to replace the C3 convolution module in YOLOv5 to obtain a training model. The dynamic snake convolution module includes at least two parallel convolution routes, each of which is provided with at least one convolution layer, and at least one convolution route is provided with a dynamic snake residual module. The convolution results of all convolution routes are aggregated and then input into the convolution layer again.
[0037] Step 3: Input the collected bird images into the training model in batches and iterate to obtain the optimal bird recognition model;
[0038] Step 4: Implement flying bird recognition through the optimal flying bird recognition model.
[0039] Further, such as Figure 2 As shown, the dynamic snake convolution module includes a first convolution layer, a second convolution layer, a dynamic snake residual module, an aggregation function module, and a third convolution layer. The first convolution layer is connected to the dynamic snake residual module, and the dynamic snake residual module is used to aggregate the output of the dynamic snake residual module with the output of the second convolution layer and input them into the third convolution layer.
[0040] Example 2:
[0041] This embodiment is further optimized on the basis of embodiment 1. The input of the first convolution layer and the second convolution layer is the same, denoted as λ; the first output of λ is obtained after passing through the first convolution layer and the dynamic snake residual module, and the first output is denoted as β1; the second output of λ is obtained after passing through the second convolution layer, and the second output is denoted as β2; the aggregation function module performs concat processing on the first output β1 and the second output β2 to fuse them into the third output β3, and the third output β3 is input to the third convolution layer for processing as the final output of the dynamic snake convolution module. Figure 3 As shown, the dynamic snake-shaped residual module includes a base convolution layer, a dynamic snake-shaped convolution layer, and an add module connected in sequence. The add module is used to add the input of the base convolution layer and the output of the dynamic snake-shaped convolution layer as the output of the dynamic snake-shaped convolution module.
[0042] The result of λ after passing through the first convolutional layer is recorded as τ. The output obtained by τ after passing through the basic convolutional layer and the dynamic snake convolutional layer in sequence is recorded as θ. After adding τ to θ, the first output β1 is obtained.
[0043] Furthermore, the convolution kernel size of the first convolutional layer is 1×1, the stride is 1, and the padding is 0.
[0044] Furthermore, the convolution kernel size of the second convolutional layer is 1×1, the stride is 1, and the padding is 0.
[0045] Furthermore, the convolution kernel size of the third convolutional layer is 1×1, the stride is 1, and the padding is 0.
[0046] Furthermore, the convolution kernel size of the base convolution layer is 1×1, the stride is 1, and the padding is 0.
[0047] The rest of this embodiment is the same as that of embodiment 1, so it will not be described again.
[0048] Example 3:
[0049] This embodiment is further optimized based on the above embodiment 1 or 2, and the step 3 specifically includes:
[0050] Step 3.1: Input the flying bird images in the training set into the training model in batches to obtain a training model set; that is, input the flying bird images in the training set into the memory model and iterative training in batches. The training model after S rounds of training is recorded as M. s , get the training model set M, M={M1,M2,...,M Q}, where S is the number of training rounds, S = {1, 2, ..., Q}, and Q is the total number of training rounds.
[0051] Step 3.2: Input the flying bird images in the validation set into the training model set and calculate the MAP index of the validation set to obtain the MAP index set. The MAP index set is recorded as O, O = O1, O2, ... O Q ;
[0052] Step 3.3: Extract the maximum index value from the MAP index set, and extract the training model corresponding to the maximum index value from the training model set as the optimal flying bird recognition model.
[0053] The rest of this embodiment is the same as that of the above-mentioned embodiment 1 or 2, and thus will not be described in detail.
[0054] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the scope of protection of the present invention.
Claims
1. A YOLOv5 bird detection method based on dynamic snake convolution, which replaces the C3 convolution module in YOLOv5 with a dynamic snake convolution module, and then inputs the collected bird images for training to obtain an optimal bird recognition model. Bird recognition is achieved through the optimal bird recognition model, characterized by: The following steps are involved: Step 1: Construct training and validation sets for bird recognition. Step 2: Build a dynamic snake convolution module. Use the dynamic snake convolution module to replace the C3 convolution module in YOLOv5 to obtain a training model. The dynamic snake convolution module includes at least two parallel convolution routes, each of which has at least one convolution layer, and at least one convolution route has a dynamic snake residual module. The convolution results of all convolution routes are aggregated and then input into the convolution layer again. Step 3: Input the collected bird images into the training model in batches and iterate to obtain the optimal bird recognition model; Step 4: Implement flying bird recognition through the optimal flying bird recognition model.
2. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 1 is characterized in that: The dynamic snake-shaped residual module includes a base convolution layer, a dynamic snake-shaped convolution layer, and an add module connected in sequence. The add module is used to add the input of the base convolution layer and the output of the dynamic snake-shaped convolution layer as the output of the dynamic snake-shaped convolution module.
3. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 2 is characterized in that: The dynamic serpentine convolution module includes a first convolution layer, a second convolution layer, a dynamic serpentine residual module, an aggregation function module, and a third convolution layer. The first convolution layer is connected to the dynamic serpentine residual module, and the dynamic serpentine residual module is used to aggregate the output of the dynamic serpentine residual module with the output of the second convolution layer and input them into the third convolution layer.
4. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 3 is characterized in that: The input of the first convolutional layer and the second convolutional layer is the same, denoted as ; After the first convolutional layer and the dynamic snake residual module, the first output is obtained. After passing through the second convolution layer, the second output is obtained; the aggregation function module performs concat processing on the first output and the second output to fuse them into a third output, and the third output is input to the third convolution layer for processing as the final output of the dynamic snake convolution module.
5. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 4 is characterized in that: The convolution kernel size of the first convolutional layer is 1×1, the stride is 1, and the padding is 0.
6. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 4 is characterized in that: The convolution kernel size of the second convolutional layer is 1×1, the stride is 1, and the padding is 0.
7. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 4, characterized in that: The convolution kernel size of the third convolutional layer is 1×1, the stride is 1, and the padding is 0.
8. The YOLOV5 bird detection method based on dynamic snake convolution according to claim 4, characterized in that: The convolution kernel size of the base convolution layer is 1×1, the stride is 1, and the padding is 0.
9. A YOLOV5 bird detection method based on dynamic snake convolution according to any one of claims 1 to 8, characterized in that: The step 3 specifically includes: Step 3.1: Input the flying bird images in the training set into the training model in batches to obtain a training model set; Step 3.2: Input the flying bird images in the validation set into the training model set and calculate the MAP index of the validation set to obtain the MAP index set; Step 3.3: Extract the maximum index value from the MAP index set, and extract the training model corresponding to the maximum index value from the training model set as the optimal flying bird recognition model.