Lightweight high-efficiency feather feature segmentation and recognition system

Abstract

A lightweight and efficient feather feature segmentation and recognition system comprises a Backbone layer backbone network module, a Neck layer fusion module, a Head layer module and a feather data preprocessing module, wherein an HPCDown downsampling module and a WTConv module are introduced into the Backbone layer backbone network module; the HPCDown downsampling module and a convolution-based cross-scale feature fusion structure CCFF structure module are introduced into the Neck layer fusion module; the Head layer module is provided with a segmentation detection Segment network. The purpose of the invention is to design a lightweight and efficient feather segmentation network according to the real-time segmentation detection task requirement of the feather production industry, so that not only is the classification detection of the feathers realized, but also the accuracy and stability of the classification detection effect of the feathers are improved; a lightweight and efficient solution is provided for a real-time segmentation detection task of a feather industrial production environment in a resource limited scene.

Description

Technical Field

[0001] The present invention designs a lightweight and efficient feather feature segmentation and recognition system, which belongs to a brand-new technology in the industrial production application field for feather feature classification and detection. Background Art

[0002] Feathers are a key component of sports products such as badminton. Feathers of different sizes and qualities will affect the flight stability and durability of badminton, which directly affects the user experience. Therefore, it is crucial to ensure the accurate classification of feathers during the production process. Traditional feather classification methods usually rely on manual inspection, which is highly subjective, labor-intensive, inefficient, costly, and easily affected by the experience, fatigue, and subjective factors of the inspectors. It is difficult to meet the needs of modern industry for high precision and high efficiency. At the same time, although automated classification based on machine vision and image processing technology can be used as a method and technology for feather feature recognition, there are still certain deficiencies in the accuracy and stability of detection when dealing with feather recognition with complex backgrounds and feather diversity.

[0003] In recent years, with the rapid development of deep learning technology, models such as convolutional neural networks (CNN) have demonstrated excellent performance in image recognition and classification tasks, and the application of deep learning technology to feather classification has become a research hotspot. Due to its adaptability and generalization ability, deep learning models can significantly improve the efficiency and reliability of feather classification and meet the needs of modern industry for high-standard quality control. Many research institutions and related companies have invested resources to develop more advanced deep learning-based classification technologies to improve product quality, reduce production costs and improve production efficiency. However, in the existing deep learning detection method technology, the network parameters are large during feature extraction, resulting in high computational complexity and slow detection speed, which makes it difficult to fully meet the real-time requirements of feather classification.

[0004] In order to improve the accuracy, stability and detection speed of feathers in classification and detection tasks, how to design a feather feature recognition deep learning model that can cope with the complex detection background of feathers, the diversity of feather types, and overcome the problems of high computational complexity and slow detection speed of current deep learning models, and build a lightweight and efficient feature segmentation method system for application in the field of industrial inspection to meet the real-time classification and detection needs of the feather industrial production environment is an urgent problem that the feather production industry needs to solve. Summary of the invention

[0005] The problem to be solved by the present invention is that, in view of the real-time segmentation and detection task requirements of the feather production industry, a lightweight and efficient feather feature segmentation and recognition system is designed, which is a lightweight and efficient HWC-YOLO feather feature segmentation and recognition network based on the HWC-YOLO network architecture. The system can realize feather classification detection, and can also improve the accuracy and stability of feather classification detection effect, providing a lightweight and efficient solution for the real-time segmentation and detection task of the feather industrial production environment under resource-constrained scenarios.

[0006] Specifically: by introducing the hybrid pooling convolution downsampling HPCDown (Hybrid Pooling ConvolutionDownsampling) network structure module, the downsampling module reduces the resolution while retaining the feather texture and background information; the convolution operation is then used to enhance the ability to capture details; in order to effectively capture the local and global features of the image, the feature extraction layer is introduced and integrated into the WTConv (Wavelet Convolutions) network structure module to perform wavelet transform decomposition on the feature map to enhance the low-frequency features of information in different frequency bands and expand the receptive field; the convolution-based cross-scale feature fusion CCFF (Convolution-based Cross-Scale Feature Fusion) network structure module is used to fuse features of different scales, thereby improving the model's adaptability to scale changes and its ability to detect feather detail textures; finally, the segmentation detection Segment network is used to locate and classify targets.

[0007] The technical solution of the present invention is: a lightweight and efficient feather feature segmentation and recognition system, comprising: a Backbone layer backbone network module introducing an HPCDown downsampling module and a WTConv module, a Neck layer fusion module introducing an HPCDown downsampling module and a convolution-based cross-scale feature fusion structure CCFF structure module, a Head layer module with a segmentation detection Segment network, and also a feather data preprocessing module;

[0008] The Backbone layer backbone network module is used to extract feather features;

[0009] The Neck layer is used to fuse feather features of different scales;

[0010] The Head layer is used to perform target classification and position regression prediction on feather feature outputs of different scales.

[0011] Furthermore, the HPCDown downsampling module is composed of a maximum pooling salient feature module, an average pooling smoothing processing module and a convolution downsampling module; the maximum pooling salient feature module is used to split the feather feature map into a first part x1 and a second part x2 for separate processing to enrich the feather features; the first part x1 of the maximum pooling salient feature module is an average pooling network, which is used to perform feather smoothing to retain background or non-extreme features; the second part x2 of the maximum pooling salient feature module is a convolution processing network, which is used to capture more spatial information while downsampling; the maximum pooling salient feature module merges the first part x1 and the second part x2 into a fused feature map along the channel dimension to process the features of feather images with complex textures and details.

[0012] Further, the WTConv module includes a WTConv_C2f module and a WTConv_C3C2f module;

[0013] The WTConv_C2f module is a C2f model based on the WTConv model, which is used to extract the features of small and medium-sized barbs and barbs in the lower layers of the Backbone layer pyramid network;

[0014] The WTConv_C3C2f module is a C3C2f model based on the WTConv model, which is used to process larger and complete feather features at high levels of the Backbone layer pyramid network.

[0015] Furthermore, the WTConv module structure is based on the C2f framework fusion WTConv_BottleNeck network;

[0016] The WTConv_C2f module is a WTConv_C2f feature extraction layer based on the C2f framework and WTConv_BottleNeck structure;

[0017] The WTConv_C3C2f module is a WTConv_C3k2 feature extraction layer based on the C2f framework and WTConv_BottleNeck structure;

[0018] The WTConv_BottleNeck network combines convolution and wavelet convolution WTConv to extract features, and then enhances the gradient flow through a residual connection to extract local features in the spatial domain.

[0019] Furthermore, the CCFF structure module includes a pyramid network FPN_Block module and a path aggregation network PAN_Block module; the FPN_Block module is used to upsample the high-level feature map of the backbone network to restore the feature map size, and then splice and fuse it with the features of the FPN_Block underlying feature map; the PAN_Block module is used to expand the receptive field of the PAN_Block underlying feature map through the HPCDown downsampling module, and then splice and fuse it with the FPN_Block high-level feature map.

[0020] Furthermore, a 1x1 convolution module (CBS module) is included between the Backbone layer and the Head layer for adjusting the input channels of the Head layer module structure, and the number of channels after adjustment is 256.

[0021] Furthermore, the feather data preprocessing module includes: a brightness changing module, a noise adding module, an image translation module, an image cropping module, and a random point adding module.

[0022] Furthermore, the Segment network is used to locate and classify feather targets.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] The present invention proposes a lightweight and efficient feather feature segmentation and recognition system based on the HWC-YOLO network structure.

[0025] The purpose of the present invention is to construct a lightweight and efficient feather segmentation algorithm to achieve classification and detection of feathers, so as to meet the real-time classification and detection requirements of the feather industrial production environment. By introducing the hybrid pooling convolution downsampling HPCDown (Hybrid Pooling Convolution Downsampling) network structure module, the downsampling module reduces the resolution while retaining the feather texture and background information; the convolution operation enhances the ability to capture details; in order to effectively capture the local and global features of the image, the feature extraction layer is introduced and integrated into the WTConv (Wavelet Convolutions) network structure module to perform wavelet transform decomposition on the feature map, enhance the low-frequency features of different frequency bands and expand the receptive field; the convolution-based cross-scale feature fusion CCFF (Convolution-based Cross-Scale Feature Fusion) network structure module is used to fuse features of different scales, improve the model's adaptability to scale changes and the ability to detect feather detail textures; finally, the segmentation detection Segment network is used to locate and classify the target.

[0026] Experimental results show that the HWC-YOLO model established in the present invention performs well in feather segmentation detection, with an average detection accuracy of 90.4%, a detection speed of 79.8FPS, a parameter amount of only 1.47M, and a computational amount of only 7.2GFLOPs. Both the speed and the model parameter amount are lightweight and efficient, solving the accuracy and stability problems of feather segmentation detection tasks, and providing a lightweight and efficient solution for real-time segmentation detection tasks in feather industrial production environments under resource-constrained scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a structural diagram of the feather segmentation network HWC-YOLO system of the present invention;

[0028] Figure 2 This is a structural diagram of the HPCDown module of the present invention;

[0029] Figure 3 This is the structural design diagram of the fusion WTConv module of the present invention;

[0030] Figure 4 The WTConv module architecture flow of the present invention;

[0031] Figure 5 This is a structural diagram of the CCFF module of the present invention.

[0032] Figure 6 It is a feather type diagram of the present invention;

[0033] Figure 7 To compare the detection effects of feather segmentation using the basic YOLO11 model and the HWC-YOLO model constructed by the present invention; DETAILED DESCRIPTION

[0034] The following is a further detailed description of the specific implementation of the present invention in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0035] Aiming at the real-time segmentation and detection task requirements of the feather production industry, the present invention designs a lightweight and efficient feather feature segmentation and recognition system, which is a lightweight and efficient feather feature segmentation and recognition network based on the HWC-YOLO network architecture. It not only realizes feather classification detection, but also improves the accuracy and stability of feather classification detection effects, providing a lightweight and efficient solution for real-time segmentation and detection tasks in feather industrial production environments under resource-constrained scenarios.

[0036] The system of the present invention introduces a hybrid pooling convolution downsampling HPCDown (Hybrid Pooling Convolution Downsampling) network structure module to reduce the resolution while retaining the feather texture and background information; the convolution operation is then used to enhance the ability to capture details; in order to effectively capture the local and global features of the image, a feature extraction layer is introduced and integrated into the WTConv (Wavelet Convolutions) network structure module to perform wavelet transform decomposition on the feature map to enhance the low-frequency features of information in different frequency bands and expand the receptive field; a convolution-based cross-scale feature fusion CCFF (Convolution-based Cross-Scale Feature Fusion) network structure module is used to fuse features of different scales, thereby improving the model's adaptability to scale changes and its ability to detect feather detail textures; finally, a segmentation detection Segment module is used to perform target positioning and classification.

[0037] The technical solution of the present invention is: a lightweight and efficient feather feature segmentation and recognition system, comprising: a Backbone layer backbone network module introducing an HPCDown downsampling module and a WTConv module, a Neck layer fusion module introducing an HPCDown downsampling module and a convolution-based cross-scale feature fusion structure CCFF structure module, a Head layer module with a segmentation detection Segment network, and also a feather data preprocessing module;

[0038] The Backbone layer backbone network module is used to extract feather features;

[0039] The Neck layer is used to fuse feather features of different scales;

[0040] The Head layer is used to perform target classification and position regression prediction on feather feature outputs of different scales.

[0041] YOLO is an end-to-end deep learning target detection model widely used in visual image detection. It can directly predict the category probability and regression box of the target from the input image. The latest ultra-lightweight version of YOLO11 has high real-time performance and lightweight network structure. The basic network structure of YOLO11 is: extracting backbone network features through the Backbone layer, fusing features of different scales in the Neck layer, and performing target classification and position regression prediction on feature outputs of different scales through the Head layer.

[0042] The feather feature segmentation network of the present invention is an improved model network based on YOLO11, i.e., the HWC-YOLO model (HWC is a combination of HPCDown, WTConv, and CCFF). The model system introduces the HPCDown downsampling module to reduce the spatial dimension, while retaining the feather texture and background information to enhance feature extraction; the WTConv_C2f module or the WTConv_C3C2f module of the WTConv network further enhances the feature extraction and fusion capabilities in the entire feather feature segmentation convolutional neural network; a 1x1 Conv convolution is added from the Backbone layer network to the Neck layer network, the number of channels is unified to 256, the network depth is reduced, and the amount of calculation is reduced; the CCFF module is used to improve the adaptability of the entire feather feature segmentation model to scale changes and fuse feather features of different scales; finally, the segmentation detection Segment network is used to locate and classify feather feature targets.

[0043] Furthermore, the HPCDown downsampling module is composed of a maximum pooling salient feature module, an average pooling smoothing processing module and a convolution downsampling module; the maximum pooling salient feature module is used to split the feather feature map into a first part x1 and a second part x2 for separate processing to enrich the feather features; the first part x1 of the maximum pooling salient feature module is an average pooling network, which is used to perform feather smoothing to retain background or non-extreme features; the second part x2 of the maximum pooling salient feature module is a convolution processing network, which is used to capture more spatial information while downsampling; the maximum pooling salient feature module merges the first part x1 and the second part x2 along the channel dimension to fuse the feature maps, which is used to integrate the high-level and low-level feature information of the feather feature segmentation network, and improve the performance of the entire model in processing feather images with complex textures and details. The structure of the HPCDown downsampling module is shown in Figure 2 shown.

[0044] Furthermore, in the design of the feather segmentation network HWC-YOLO, in order to improve the segmentation network's ability to extract feather features, the WTConv module is integrated in the Backbone layer network and the feature extraction layer of the Neck layer to perform feature fusion and feature enhancement.

[0045] The WTConv module includes a WTConv_C2f module and a WTConv_C3C2f module. Figure 3 shown.

[0046] In order to better extract feather features, the WTConv_C2f module is a C2f model based on the WTConv model, which is used to extract small-sized and medium-sized feather branches and feather roots in the lower layer of the Backbone layer pyramid network;

[0047] In order to better extract feather features, the WTConv_C3C2f module is a C3C2f model based on the WTConv model, which is used to process larger and complete feather features at high levels of the Backbone layer pyramid network.

[0048] Furthermore, the WTConv module is based on the C2f framework fusion WTConv_BottleNeck network;

[0049] The WTConv_C2f module is a WTConv_C2f feature extraction layer network based on the C2f framework fused with the WTConv_BottleNeck network. The network module has low computational complexity and has local feature extraction capabilities through residual fusion. It is suitable for feather feature segmentation and recognition pyramid network bottom-level feature extraction, ensuring that the segmentation model is lightweight and efficient.

[0050] The WTConv_C3C2f module is a WTConv_C3k2 feature extraction layer network based on the C2f framework fused with the WTConv_BottleNeck network. The C3 module of this network module has multi-branch feature fusion and enhanced diversity, which is suitable for feather feature segmentation and recognition pyramid structure high-level feature extraction, and the learning of feather details and semantic information is more comprehensive.

[0051] The WTConv_BottleNeck network combines convolution and wavelet convolution WTConv to extract features, and then enhances the gradient flow through a residual connection to extract local features in the spatial domain. Compared with the BottleNeck basic model, which mainly extracts local features in the spatial domain and consists of two convolutions, the WTConv_BottleNeck network combines spatial and frequency domain information to capture multi-scale features. The WTConv_BottleNeck network architecture is shown in Figure 3 shown.

[0052] The WTConv module of the present invention is an improved wavelet convolution module that combines traditional convolution operations and wavelet transforms. This module can improve the expressive power of the model, especially when processing high-frequency and low-frequency information of an image, and enhance the ability to extract features. Wavelet transform can perform multi-scale analysis on an image and decompose information of different frequency bands. The corresponding inverse filter is used for the feature map after wavelet transform, and the original feature space is restored through a transposed convolution operation. This process effectively captures the local and global features of the feather image, and enhances the expressive power of the model, especially in the high-frequency details of the feather. The WTConv network architecture is shown in Figure 4 shown.

[0053] Furthermore, in order to reduce the computational complexity of the HWC-YOLO model and improve the ability to fuse features of different scales, a convolution-based cross-scale feature fusion structure CCFF module is introduced in the HWC-YOLO network of the present invention. The CCFF structure module is mainly designed for the Neck feature fusion layer of the HWC-YOLO model. The high-level information and the bottom-level information are spliced ​​through the Concat network through convolution, and then the cross-scale feature fusion structure is performed by the feature fusion layer designed by the WTConv module. The CCFF module is shown in Figure 5 shown.

[0054] The CCFF structure module includes a pyramid network FPN_Block module and a path aggregation network PAN_Block module; the FPN_Block module is used to upsample the high-level feature map of the backbone network to restore the feature map size, and then splice and fuse it with the features of the FPN_Block bottom-level feature map; the PAN_Block module is used to expand the receptive field of the PAN_Block bottom-level feature map through the HPCDown downsampling module, and then splice and fuse it with the FPN_Block high-level feature map.

[0055] Furthermore, a 1x1 convolution module (CBS module) is included between the Backbone layer and the Head layer for adjusting the input channel of the Head layer module structure. The number of channels after adjustment is 256. Figure 1 As shown. This step greatly reduces the parameter calculation of the HWC-YOLO model, while increasing the nonlinear expression ability of the entire network, helping the network learn more complex feather features. After the CBS module, the FPN_Block and PAN_Block designed by the CCFF structure greatly improve the ability of the deep learning model to process multi-scale information, and fully utilize the features of different levels through cross-scale feature fusion, especially in visual tasks such as feather detection that require processing high-resolution details and low-resolution semantic information at the same time. It can better identify and distinguish feathers of different types and sizes, and improve the overall detection performance of the model.

[0056] Furthermore, in order to more effectively improve the robustness and generalization ability of the feather segmentation detection model, fully learn the characteristics of the feather shaft, barb and root, and perform systematic and diversified data enhancement on the collected feather data, a feather data preprocessing module is set up.

[0057] The feather data preprocessing module includes five transformation modules, which are: brightness change module, noise adding module, image translation module, image cropping module and random point adding module.

[0058] Change the brightness module to adjust the brightness of the image to simulate the feather shooting effects under different ambient light conditions, and more realistically reflect the lighting changes in nature and artificial environments;

[0059] A noise adding module is used to add noise to the feather image to simulate the sensor noise or environmental noise during the shooting process;

[0060] Image translation module, which simulates shooting at different positions by translating feather images horizontally or vertically;

[0061] An image cropping module that randomly crops different parts of the image to simulate the situation where the feather is not completely in the field of view;

[0062] The Add Random Points module adds random points to the feather image to simulate the effects of dust, water droplets or other visual distractions.

[0063] Furthermore, the Segment network is used to locate and classify feather targets.

[0064] Effects of the embodiment:

[0065] In order to develop the effect of the present invention, the experiment uses an infrared camera to capture the shape and size of each feather before it passes through the assembly line. The image resolution is: 2448 pixels x 2048 pixels. The feathers are classified by the detection system to distinguish the different sizes and types of feathers on the control assembly line. Figure 6 shown.

[0066] Experimental example 1: ablation experiment.

[0067] In order to verify the contribution of each design module to the feather feature segmentation detection performance, an ablation experiment was conducted on the recognition network constructed by the present invention, and the results are shown in Table 1. Among them, recognition model 1 uses the baseline YOLO11 model as the experimental baseline, and recognition models 2 to 8 are ablation experiments on the permutation and combination of each design module, and recognition model 8 is the HWC-YOLO model of the present invention. The ablation experiment items include: precision P (Precision), recall rate R (Recall), average detection accuracy mean mAP (meanAverage Precision) to evaluate model accuracy, model parameter quantity (Params), model calculation quantity (GFLOPs) to evaluate the computational complexity of the model, and detection frame rate FPS (Frames Per Second) to evaluate the real-time performance of the model.

[0068] Table 1 Ablation experiment results

[0069]

[0070] From the analysis of Table 1, we can see that: in terms of the accuracy and average detection accuracy of feather detection, both the introduction of a single module and the introduction of a combination of modules are better than the baseline YOLO11 model, and the average detection accuracy of the HWC-YOLO model is improved to 90.4%; in terms of the number of model parameters and the amount of calculation, the HPCDown downsampling module alone, the feature extraction structure of the WTConv module fusion and the CCFF structure, the number of parameters and the amount of calculation are reduced by more than 10% compared with the baseline model, and the combination of the two modules is reduced by more than 18%. The number of parameters and the amount of calculation of the HWC-YOLO model are reduced by 48.1% and 31.4% respectively compared with YOLO11, further reducing the complexity and number of parameters of the model; in terms of the detection speed of the model, the introduction of individual modules is not as high as the baseline model, but the detection speed of the combination of modules is basically higher than the baseline model.

[0071] Based on the above analysis, the HWC-YOLO model has lightweight parameters, efficient detection speed, and high detection accuracy.

[0072] Experimental Example 2: Comparative experiment on the recognition effects of different YOLO models.

[0073] The performance of advanced deep learning segmentation detection algorithms such as YOLOv5, YOLOv8 and YOLOv9 is compared with the HWC-YOLO network model algorithm proposed in this paper, and the results are shown in Table 2. The model parameters (Params) and model calculation amount (GFLOPs) are used to evaluate the computational complexity of the model, the average detection accuracy mAP (mean Average Precision) is used to evaluate the model accuracy, and the detection frame rate FPS (Frames Per Second) is used to evaluate the real-time performance of the model as comparative experimental items.

[0074] Table 2 Comparative experiments on recognition effects of different YOLO models

[0075]

[0076] By comparison, it can be seen that when the experimental environment remains unchanged and 300 rounds of training are uniformly performed, the YOLO11n model is compared with the lightweight YOLOv5n and YOLOv8n models. Although the former has relatively low parameter and computational complexity, its detection accuracy mAP is about 5% lower. When the latter have similar detection accuracy mAP, the YOLO11n model has lower computational complexity and parameter complexity and higher detection speed.

[0077] Compared with the YOLO11n model, the HWC-YOLO model has higher parameters and detection efficiency than the YOLO11n model. Compared with the YOLOv9e, although the detection accuracy mAP is slightly lower, the parameter amount, computational complexity and detection efficiency are all better.

[0078] In order to verify the detection effect of the improved HWC-YOLO model, a group of left-side feather-like images with consistent directions and similar shapes were selected to conduct experiments on YOLO11n model recognition and HWC-YOLO model recognition to compare the results. Figure 7 shown.

[0079] By comparison, it can be seen that both algorithms can segment and detect feathers of various categories, but the confidence of the HWC-YOLO model is higher than that of YOLO11. Both the global features of large feathers and the local features of small feathers can be effectively improved, verifying the effectiveness of the HPCDown, WTConv, and CCFF improvement strategies for feather detection.

[0080] Based on the above comparative experiments, the HWC-YOLO model system of the present invention performs well in feather segmentation detection. The proposed HWC-YOLO model system achieves a good balance between model lightweight and detection performance, and achieves the purpose of lightweight and high efficiency in terms of speed and model parameter quantity. While maintaining high detection accuracy, it significantly reduces the computational complexity and has strong real-time performance, which solves the accuracy and stability problems of feather segmentation detection tasks, and provides an efficient solution for segmentation detection tasks in resource-constrained scenarios.

[0081] It should be noted that, for those skilled in the art, several equivalent variations, substitutions and improvements can be made without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A lightweight and efficient feather feature segmentation and recognition system, characterized in that: include: The Backbone layer backbone network module that introduces the HPCDown downsampling module and the WTConv module, the Neck layer fusion module that introduces the HPCDown downsampling module and the convolution-based cross-scale feature fusion structure CCFF structure module, the Head layer module with the segmentation detection Segment network, and also includes a feather data preprocessing module; The Backbone layer backbone network module is used to extract feather features; The Neck layer is used to fuse feather features of different scales; The Head layer is used to perform target classification and position regression prediction on feather feature outputs of different scales.

2. A lightweight and efficient feather feature segmentation and recognition system according to claim 1, characterized in that The HPCDown downsampling module is composed of a maximum pooling salient feature module, an average pooling smoothing processing module and a convolution downsampling module; the maximum pooling salient feature module is used to split the feather feature map into a first part x1 and a second part x1 for separate processing to enrich the feather features; the first part x1 of the maximum pooling salient feature module is an average pooling network, which is used to perform feather smoothing processing to retain background or non-extreme features; The second part x2 of the maximum pooling salient feature module is a convolution processing network, which is used to capture more spatial information while downsampling; the maximum pooling salient feature module merges the first part x1 and the second part x2 along the channel dimension to fuse the feature map and process feather images with complex textures and details.

3. A lightweight and efficient feather feature segmentation and recognition system according to claim 1, characterized in that The WTConv module includes a WTConv_C2f module and a WTConv_C3C2f module; The WTConv_C2f module is a C2f model based on the WTConv model, which is used to extract the features of small and medium-sized barbs and barbs in the lower layers of the Backbone layer pyramid network; The WTConv_C3C2f module is a C3C2f model based on the WTConv model, which is used to process larger and complete feather features at high levels of the Backbone layer pyramid network.

4. A lightweight and efficient feather feature segmentation and recognition system according to claim 3, characterized in that The WTConv module is used for feature fusion and feature enhancement; The WTConv module structure is based on the C2f framework fusion WTConv_BottleNeck network; The WTConv_C2f module is a WTConv_C2f feature extraction layer based on the C2f framework and WTConv_BottleNeck structure; The WTConv_C3C2f module is a WTConv_C3k2 feature extraction layer based on the C2f framework and WTConv_BottleNeck structure; The WTConv_BottleNeck network combines convolution and wavelet convolution WTConv to extract features, and then enhances the gradient flow through a residual connection to extract local features in the spatial domain.

5. A lightweight and efficient feather feature segmentation and recognition system according to claim 1, characterized in that The CCFF structure module includes a pyramid network FPN_Block module and a path aggregation network PAN_Block module; the FPN_Block module is used to upsample the high-level feature map of the backbone network to restore the feature map size, and then splice and fuse it with the features of the FPN_Block bottom-level feature map; the PAN_Block module is used to expand the receptive field of the PAN_Block bottom-level feature map through the HPCDown downsampling module, and then splice and fuse it with the FPN_Block high-level feature map.

6. A lightweight and efficient feather feature segmentation and recognition system according to claim 1, characterized in that A 1x1 convolution module is also included between the Backbone layer and the Head layer, which is used to adjust the input channels of the Head layer module structure, and the number of channels after adjustment is 256.

7. A lightweight and efficient feather feature segmentation and recognition system according to claim 1, characterized in that The feather data preprocessing module includes: a brightness changing module, a noise adding module, an image translation module, an image cropping module, and a random point adding module.

Images