A feather shape classification method and system based on deep learning

Abstract

The application discloses a kind of feather shape classification method and system based on deep learning, method includes: obtaining the shape image of feather to be classified;The shape image is preprocessed, and feather image is obtained;Feather segmentation model is used to carry out segmentation processing to feather image, and feather morphology image and feather morphology image are obtained;Feather classification model is used to carry out morphology grade classification to feather morphology image, and first classification score value is obtained;Feather morphology image is converted to target HSV color space, and then second classification score value is obtained by color threshold processing;First classification score value and second classification score value are weighted calculation, and the total classification score value of feather to be classified is obtained.The application comprehensively considers various shape factors of feather, and then the feather shape classification can be accurately and comprehensively carried out, and can be widely applied in computer technology field.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for classifying feather shapes based on deep learning. Background Technology

[0002] The badminton shuttlecock appearance defect classification algorithm is a technology used to detect and classify appearance defects in badminton products. During the production and quality inspection of badminton shuttlecocks, appearance defects can lead to a decline in product quality, affect user experience, and even result in product rejection. Therefore, defect classification algorithms have significant application value in the badminton industry.

[0003] Currently, the market commonly uses traditional visual algorithms to solve defect problems. This method, based on image processing and machine learning techniques, is simple to implement, has a relatively fast computation speed, and does not require a large amount of data for training. It can achieve good results on a small sample set and is suitable for situations with limited data. However, existing methods have limited effectiveness in complex scenes and perform poorly in cases with diverse defects and significant changes in lighting. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a deep learning-based method and system for feather shape classification, which can accurately classify feather shapes.

[0005] On one hand, embodiments of the present invention provide a deep learning-based feather shape classification method, including:

[0006] Obtain the external image of the feathers to be classified;

[0007] The shape image is preprocessed to obtain a feather image;

[0008] The feather image is segmented using a feather segmentation model to obtain a feather shape image and a feather root shape image.

[0009] The feather segmentation model is obtained by training the Unet semantic segmentation network on feather images labeled with feather morphology and feather root morphology.

[0010] Using a feather classification model, the feather morphology images are classified into morphological levels to obtain a first classification score.

[0011] The feather classification model is constructed based on a residual neural network. The output layer of the residual neural network is a multi-level classifier, and each level of the classifier corresponds to a different morphological level.

[0012] The feather root image is converted to the target HSV color space, and then the second classification score is obtained through color thresholding.

[0013] The first classification score and the second classification score are weighted and calculated to obtain the total classification score of the feather to be classified.

[0014] Optionally, the shape image is preprocessed, including at least one of the following:

[0015] The shape image is subjected to image enhancement processing;

[0016] The shape image is denoised.

[0017] The shape image is cropped based on a preset size.

[0018] Optionally, the method further includes:

[0019] The feather segmentation model is created based on feather images labeled with feather morphology tags and feather root morphology tags.

[0020] Optionally, based on feather images labeled with feather morphology tags and feather root morphology tags, a feather segmentation model is created, including:

[0021] A training set is determined based on feather images labeled with feather morphology and feather root morphology; then, data augmentation is performed on the training set.

[0022] The Unet semantic segmentation network is set up, and the Unet semantic segmentation network is trained for classification using the augmented training set. Based on the training results, the Unet semantic segmentation network is adjusted to obtain the feather segmentation model.

[0023] Optionally, the training set may be augmented by at least one of the following:

[0024] The images in the training set are randomly cropped to generate new training samples which are then added to the training set.

[0025] The images in the training set are randomly flipped to generate new training samples which are then added to the training set.

[0026] The images in the training set are randomly scaled to generate new training samples which are then added to the training set.

[0027] Optionally, the method further includes:

[0028] Using the PyTorch deep learning framework, a pre-trained ResNet-152 model is loaded, and the output layer of the ResNet-152 model is replaced with a multi-level classifier to obtain the feather classification model.

[0029] Optionally, the feather root morphology image is converted to the target HSV color space, and then a second classification score is obtained through color thresholding, including:

[0030] The feather shape image is converted to the target HSV color space to obtain the converted image;

[0031] Based on a preset color threshold, pixel extraction is performed on the converted image to obtain a binary image;

[0032] The binary image is filtered by area to obtain a filtered image;

[0033] The selected images are divided into color levels by preset different color threshold levels to obtain the second category score.

[0034] On the other hand, embodiments of the present invention provide a deep learning-based feather shape classification system, comprising:

[0035] The first module is used to acquire images of the shape of the feathers to be classified;

[0036] The second module is used to preprocess the shape image to obtain a feather image;

[0037] The third module is used to segment the feather image using a feather segmentation model to obtain a feather shape image and a feather root shape image.

[0038] The feather segmentation model is obtained by training the Unet semantic segmentation network on feather images labeled with feather morphology and feather root morphology.

[0039] The fourth module is used to classify the feather morphology image into morphological levels using a feather classification model to obtain a first classification score.

[0040] The feather classification model is constructed based on a residual neural network. The output layer of the residual neural network is a multi-level classifier, and each level of the classifier corresponds to a different morphological level.

[0041] The fifth module is used to convert the feather root morphology image to the target HSV color space, and then obtain the second classification score value through color thresholding.

[0042] The sixth module is used to perform a weighted calculation on the first classification score and the second classification score to obtain the total classification score of the feathers to be classified.

[0043] Optionally, the system also includes:

[0044] The seventh module is used to create the feather segmentation model based on feather images labeled with feather morphology tags and feather root morphology tags.

[0045] Optionally, the system also includes:

[0046] The eighth module is used to load a pre-trained ResNet-152 model using the PyTorch deep learning framework, and replace the output layer of the ResNet-152 model with a multi-level classifier to obtain the feather classification model.

[0047] On the other hand, embodiments of the present invention provide an electronic device, including a processor and a memory;

[0048] Memory is used to store programs;

[0049] The processor executes the program as described above.

[0050] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.

[0051] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0052] This invention first acquires an image of the shape of a feather to be classified; preprocesses the shape image to obtain a feather image; uses a feather segmentation model to segment the feather image to obtain a feather morphology image and a feather root morphology image; wherein, the feather segmentation model is trained on a Unet semantic segmentation network using feather images labeled with feather morphology and feather root morphology; uses a feather classification model to classify the feather morphology image into morphological levels to obtain a first classification score; wherein, the feather classification model is constructed based on a residual neural network, the output layer of which is a multi-level classifier, each level of which corresponds to a different morphology level; the feather root morphology image is converted to a target HSV color space, and then processed by color thresholding to obtain a second classification score; the first classification score and the second classification score are weighted and calculated to obtain the total classification score of the feather to be classified. This invention separates the feather shape and the feather root shape, and then obtains corresponding classification scores for the feather shape grade and the feather root color grade respectively. Finally, the two scores are weighted and calculated, taking into account various shape factors of the feather, so as to comprehensively and accurately classify the feather shape. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 A flowchart illustrating a deep learning-based feather shape classification method provided in an embodiment of the present invention;

[0055] Figure 2 This is a schematic diagram of the structure of the Unet semantic segmentation network provided in an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram illustrating the segmentation of the feather root outline provided in an embodiment of the present invention;

[0057] Figure 4 A schematic diagram illustrating an example of an augmented image of the feather shape provided in an embodiment of the present invention;

[0058] Figure 5 This is a schematic diagram of the residual block structure provided in an embodiment of the present invention;

[0059] Figure 6A schematic diagram illustrating the overall process of the deep learning-based feather shape classification method provided in this embodiment of the invention;

[0060] Figure 7 A schematic diagram of the overall architecture of feather shape classification based on deep learning provided in an embodiment of the present invention;

[0061] Figure 8 A schematic diagram of a deep learning-based feather shape classification system provided in this embodiment of the invention;

[0062] Figure 9 This is a schematic diagram of the frame of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] First, it should be noted that, for the sake of understanding the technical solution of this invention, the technical terms that may appear in this invention will be explained:

[0065] On the one hand, such as Figure 1 As shown, an embodiment of the present invention provides a deep learning-based feather shape classification method, including:

[0066] S100. Obtain the shape image of the feathers to be classified;

[0067] In some specific embodiments, image data of the shape of a badminton shuttlecock can be acquired in real time using an industrial camera.

[0068] S200: Preprocess the shape image to obtain a feather image;

[0069] It should be noted that in some embodiments, the preprocessing of the shape image includes at least one of the following: image enhancement processing of the shape image; noise reduction processing of the shape image; and cropping processing of the shape image based on a preset size.

[0070] In some specific embodiments, the image needs to be preprocessed, including image enhancement (e.g., frequency domain and spatial domain methods), denoising (which can be achieved through mean filters, adaptive Wiener filters, morphological noise filters, or median filters), cropping, and other operations to improve the subsequent image processing results.

[0071] S300. Using a feather segmentation model, the feather image is segmented to obtain a feather shape image and a feather root shape image.

[0072] The feather segmentation model is obtained by training the Unet semantic segmentation network on feather images labeled with feather morphology and feather root morphology.

[0073] It should be noted that, in some embodiments, the method may further include: creating the feather segmentation model based on feather images labeled with feather morphology tags and feather root morphology tags.

[0074] In some embodiments, creating a feather segmentation model based on feather images labeled with feather morphology and feather root morphology may include: determining a training set based on the feather images labeled with feather morphology and feather root morphology; then augmenting the training set; setting up a Unet semantic segmentation network; training the Unet semantic segmentation network for classification using the augmented training set; and adjusting the Unet semantic segmentation network based on the training results to obtain the feather segmentation model.

[0075] In some embodiments, data augmentation of the training set includes at least one of the following: randomly cropping images in the training set to generate new training samples and adding them to the training set; randomly flipping images in the training set to generate new training samples and adding them to the training set; or randomly scaling images in the training set to generate new training samples and adding them to the training set.

[0076] In some specific embodiments, a method based on the Unet semantic segmentation network can be used to segment feathers and feather roots. Unet is a nine-layer deep U-shaped network structure proposed by Olaf Ronneberger. Its suitability for small datasets and other advantages have led to its widespread application in the medical field. The network is divided into downsampling and upsampling. Downsampling is the traditional feature extraction process. Upsampling also has many characteristics, such as the number of channels fused with the downsampling and feature extraction parts. Figure 2 As shown, the first part of UNet is the backbone feature extraction section, which obtains a series of feature layers. Similar to VGG, the backbone feature extraction section of UNet uses stacked convolutions and max pooling. Using the backbone feature extraction section, we obtain five preliminary effective feature layers. In the second step, these five effective feature layers are used for feature fusion. The second part is the enhanced feature extraction section. The experiment uses the five preliminary effective feature layers obtained from the backbone section for upsampling and feature fusion to obtain a final effective feature layer that integrates all features. The third part is the prediction section, which uses the final effective feature layer to classify each feature point, essentially classifying each pixel.

[0077] It's important to note that data annotation is a crucial step in training object detection models; labeling the dataset appropriately can improve the model's accuracy and robustness. In the feather root segmentation task, the acquired images need to be annotated using polygons to mark the boundaries between feathers and root roots. The annotated file is in JSON format, and the polygon position information is contained within the annotation file. The annotated image is shown below. Figure 3 As shown.

[0078] After data collection, data augmentation is necessary. This involves a series of random transformations or deformations of the images to generate new datasets, increasing the sample size and diversity, thereby improving the model's training performance. Common data augmentation methods include mirroring, rotation, scaling, translation, random cropping, adding noise, and color transformation. The appropriate data augmentation method should be selected based on the specific situation. For example, random cropping and mirroring can increase dataset diversity, rotation and scaling can increase the model's adaptability to different angles and sizes, and adding noise and color transformation can increase the model's robustness.

[0079] Due to the limited dataset used, OpenCV is required for data augmentation during the training process. OpenCV (Open Source Computer Vision Library) is an open-source computer vision and machine learning library that provides a series of functions and tools for processing images and videos. OpenCV has become a cross-platform computer vision technology standard. It provides multiple modules, including a core module, image I / O and display module, image processing module, feature detection and description module, computer vision module, machine learning module, and deep learning module. These modules offer different functions and tools and are widely used in image and video processing, robotics, autonomous driving, medical imaging, and other fields. The OpenCV data augmentation techniques used in this embodiment of the invention include the following:

[0080] Random cropping: Randomly cropping a portion of the original image to generate new training samples. In practical applications, images are usually cropped according to a certain ratio, such as cropping the original image by a ratio of 0.9.

[0081] Random Flip: Randomly flips the original image to generate new training samples. The flip direction is usually randomly generated and can be horizontal, vertical, or random.

[0082] Random scaling: Randomly scaling the original image to generate new training samples. The scaling ratio is usually randomly generated within a certain range, such as random scaling between 0.8 and 1.2.

[0083] By using these techniques, the size of the training set can be increased, and the robustness and generalization performance of the model can be improved, such as... Figure 4 This is an example of an augmented image.

[0084] In some specific embodiments, the training of the Unet semantic segmentation model includes the following process:

[0085] Input and output: The input is a badminton shuttlecock shape image, and the output is a binarized segmented image, that is, classifying each pixel to determine whether it belongs to a feather.

[0086] Network Structure: The main framework adopts the standard U-Net encoder-decoder structure. The backbone feature extraction network uses ResNet18. In the ResNet18 backbone network, the residual block is the core component of the network, which can effectively alleviate the gradient vanishing problem. In this structure, the shortcut connections between the input and output of each residual block enable the model to better capture details and features in the image, thereby improving the model's performance and stability. Simultaneously, a model pre-trained on the ImageNet dataset is used as the encoder weights, and then fine-tuned on the target task to accelerate training and improve performance.

[0087] Loss function: A pixel-level binary cross-entropy loss function is used, with a weighting factor added to balance the number of positive and negative samples.

[0088] During training, an optimizer is used to update network parameters using gradient descent to minimize the loss function. After training, the trained U-Net model is used to predict new feather images and segment the corresponding binarized feather images.

[0089] S400. Using a feather classification model, the feather morphology image is classified into morphological levels to obtain a first classification score.

[0090] The feather classification model is constructed based on a residual neural network. The output layer of the residual neural network is a multi-level classifier, and each level of the classifier corresponds to a different morphological level.

[0091] It should be noted that in some embodiments, the method may further include: using the PyTorch deep learning framework, loading a pre-trained ResNet-152 model, and replacing the output layer of the ResNet-152 model with a multi-level classifier to obtain the feather classification model.

[0092] In some specific embodiments, the segmented feather morphology images can be fed into a ResNet152 image classification neural network for model training to complete the classification of feather defect levels. ResNet152 is one of the ResNet (Residual Neural Network) series models, proposed by the Microsoft Research team in 2015. It is a variant of deep convolutional neural networks, an extension of ResNet-50 and ResNet-101, with a deeper and wider network structure. The main innovation of ResNet152 is the introduction of residual blocks, through residual connections, such as... Figure 5 As shown, the input and output features are added together, making it easier for the network to learn the identity mapping. This design solves the vanishing and exploding gradient problems, allowing the network to be trained deeper and thus obtain stronger feature representation capabilities. The overall structure of ResNet-152 contains multiple deep residual blocks, each containing multiple convolutional layers, batch normalization layers, and ReLU activation functions. Cross-layer links are established between the input and output of each residual block, realizing residual connections. In addition, ResNet-152 uses global average pooling layers to transform feature maps into global feature vectors, reducing the number of parameters and improving computational efficiency.

[0093] In some specific embodiments, the training of the ResNet152 image classification model includes the following process:

[0094] Data preprocessing: First, the segmented feather morphology images are preprocessed, including image resizing, normalization, and enhancement. This converts the images into a format suitable for input to the ResNet-152 model.

[0095] Dataset construction: The preprocessed images are divided into training and validation sets for model training and evaluation.

[0096] Loading the ResNet-152 model: Using the PyTorch deep learning framework, load the pre-trained ResNet-152 model.

[0097] Replace the classification layer: Replace the last classifier layer of the ResNet-152 model with a 5-layer classifier, representing five levels of defect. Since the category of the feather morphology image may differ from the category of the original image classification task, the output layer needs to be redefined.

[0098] Model training: The ResNet-152 model, loaded with segmented feather images, was trained on the training set to optimize the classifier parameters.

[0099] Model inference: The feather image is subjected to defect level inference to obtain the defect level and the score value of the region of that level. The score is used for subsequent weighted calculation.

[0100] S500: Convert the feather root morphology image to the target HSV color space, and then obtain the second classification score value through color thresholding.

[0101] It should be noted that, in some embodiments, step S500 may include: converting the feather morphology image to the target HSV color space to obtain a converted image; extracting pixels from the converted image based on a preset color threshold to obtain a binary image; performing area filtering on the binary image to obtain a filtered image; and dividing the filtered image into color levels using preset different color threshold levels to obtain a second classification score.

[0102] In some specific embodiments, color levels can be classified based on the shape of the feather root. The feather root color threshold level classification includes the following process:

[0103] The segmented feather image is converted to the specified HSV (Hue, Saturation, Lightness) color space. The HSV color space can separate color information and lightness information, which facilitates subsequent color thresholding processing.

[0104] Color thresholding: The converted image is subjected to color thresholding. By setting an appropriate threshold, pixels that meet the criteria are extracted from the image to obtain a binary image of the feather.

[0105] Area filtering: Since defects may only occupy a small part of the entire feather root image, it is necessary to perform area filtering on the binary image to remove excessively small connected regions and retain only the feather root regions that may contain defects.

[0106] Color Grading: The feather root area is classified according to the color threshold. Based on different threshold settings, the feather root can be divided into different color grades, such as light yellow, normal, dark yellow, etc. The color grade and the score of the grade area are obtained, and the score is used for subsequent weighted calculation.

[0107] S600. The first classification score and the second classification score are weighted and calculated to obtain the total classification score of the feathers to be classified.

[0108] In some specific embodiments, the overall defect classification of the badminton shuttlecock is achieved by weighting the feather defect score (i.e., the first category score) and the feather root color threshold score (i.e., the second category score).

[0109] The specific method for weighted calculation of feather defect score and feather root color threshold score can be adjusted according to actual needs and data conditions. The following is the weighted calculation method:

[0110] The feather defect score is called score_defect, which ranges from 0 to 1 and represents the degree of defect. The closer it is to 1, the more serious the defect.

[0111] The feather color threshold score is score_color, which ranges from 0 to 1. It represents the color level, and the closer it is to 1, the closer the color is to the ideal state.

[0112] We define the weight of the defect score as w_defect and the weight of the color threshold score as w_color, with a sum of 1. Based on these settings, the weighted calculation formula is as follows:

[0113] Final_score=w_defenct×score_defect+w_color×score_color

[0114] The values ​​w_defect and w_color can be set according to actual needs, indicating the degree of importance given to defect scores and color scores in the overall rating. This requirement necessitates a higher weighting for defect scores to reflect the overall classification of defects in badminton shuttlecocks.

[0115] To illustrate the technical principles of the embodiments of the present invention in detail, the present invention will be further described below with reference to the accompanying drawings and some specific embodiments. It should be understood that the following is an explanation of the present invention and should not be regarded as a limitation of the present invention.

[0116] In some specific embodiments, such as Figure 6 As shown, image data of the badminton shuttlecock's shape is acquired in real time using an industrial camera. The images require preprocessing, including image enhancement (e.g., frequency domain and spatial domain methods), denoising (using mean filters, adaptive Wiener filters, morphological noise filters, or median filters), and cropping, to improve subsequent image processing results. First, a Unet-based semantic segmentation network is used to segment the feathers and feather root. The segmented feather shape images are then fed into a ResNet152 image classification neural network for model training, completing the classification of feather defect levels. Feather root shapes are classified by color level, achieving overall defect classification of the badminton shuttlecock's shape. Through this software design approach, a deep learning-based badminton shuttlecock shape defect level classification algorithm can achieve defect level sorting for badminton shuttlecocks, improving work efficiency.

[0117] In some specific embodiments, such as Figure 7As shown, the overall architecture of this invention is as follows: First, various image datasets of badminton shuttlecock shapes are collected. Experienced defect sorting personnel classify the feather shapes according to their levels, and data augmentation is performed to complete the data processing. In the model training experiment, the processed dataset is uploaded to the training server. First, a method based on the Unet semantic segmentation network is used to segment the feathers and feather root. The segmented feather shape images are loaded into a ResNet152 image classification neural network for model training to obtain the defect level of each feather. After model evaluation, the model is deployed online to the inference server. The feather root shape is classified by color level to achieve the overall defect classification of the badminton shuttlecock shape. User-uploaded images are stored using an encryption protocol, and the stored image data is de-identified before being used for model training again. This solves the data encryption problem and also helps the model to continue to iterate and update.

[0118] In some specific embodiments, the aforementioned models can be converted into a format suitable for edge computing boxes. The PaddlePaddle model is converted into an RKNN model. The model is deployed to the edge computing box. The camera input stream is imported into the edge computing box. This can be achieved by directly connecting the camera to the edge computing box, or by transmitting the camera input stream to the edge computing box via a network. Inference is performed on the model on the edge computing box to detect the defects in the badminton shuttlecock in real time. Corresponding control operations are performed based on the classification results.

[0119] In summary, based on the limitations of existing technologies, this invention proposes a deep learning-based defect classification algorithm for feather shapes. This algorithm automatically learns image features using a deep convolutional neural network (CNN) to achieve efficient detection and classification of defects in the appearance of badminton shuttlecocks. Compared with traditional methods, this algorithm has better adaptability and generalization ability, and can effectively handle complex defect scenarios and large-scale datasets. Simultaneously, it enables real-time data transmission and storage, facilitating automated and intelligent production. This can improve product quality, reduce manual quality inspection costs, increase production efficiency, and also help improve user experience and enhance brand image. Compared with existing technologies, this invention has at least the following beneficial effects:

[0120] High accuracy: Deep learning models possess powerful feature learning capabilities, enabling them to learn complex image features and achieve highly accurate defect classification. Compared to traditional rule-based or manual feature extraction methods, deep learning can better capture subtle features, thus improving classification accuracy.

[0121] Automation and efficiency: Deep learning algorithms can achieve end-to-end automated processing, eliminating the need for manual feature or rule design and saving significant manual labor and time costs. Furthermore, deep learning models have fast inference speeds, enabling them to quickly process large amounts of image data.

[0122] Diverse Defect Classification: Deep learning models can learn the features of multiple defect categories, thus achieving diverse defect classification. This provides a more comprehensive understanding of badminton shuttlecock shape issues, helping to improve product quality and manufacturing process control.

[0123] Transferability: Based on pre-training, deep learning models can adapt to different datasets and scenarios through transfer learning, exhibiting strong generalization capabilities. This means that even on new production lines or product models, a defect classification system can be quickly built with minimal adjustments and data collection.

[0124] Real-time monitoring and early warning: Deep learning-based defect classification algorithms can achieve real-time image processing and monitoring, promptly detect appearance defects, reduce the number of defective products, and improve product quality and production efficiency.

[0125] Data-driven decision-making: Deep learning models can learn from massive amounts of data, thus providing data-driven decision support. By analyzing defect classification results and statistical data, production processes can be optimized, product design improved, and the overall quality of badminton shuttlecocks enhanced.

[0126] The deep learning-based badminton shuttlecock shape defect classification algorithm will bring many beneficial effects to badminton shuttlecock production and quality control, thereby improving product quality, reducing production costs, and providing enterprises with more effective quality management methods.

[0127] On the other hand, such as Figure 8As shown, an embodiment of the present invention provides a deep learning-based feather shape classification system 700, comprising: a first module for acquiring a shape image of a feather to be classified; a second module for preprocessing the shape image to obtain a feather image; a third module for segmenting the feather image using a feather segmentation model to obtain a feather morphology image and a feather root morphology image; wherein the feather segmentation model is trained on a Unet semantic segmentation network using feather images labeled with feather morphology and feather root morphology labels; a fourth module for classifying the feather morphology image into morphological levels using a feather classification model to obtain a first classification score; wherein the feather classification model is constructed based on a residual neural network, the output layer of the residual neural network being a multi-level classifier, each level of the classifier corresponding to a different morphological level; a fifth module for converting the feather root morphology image to a target HSV color space, and then obtaining a second classification score through color thresholding; and a sixth module for weighted calculation of the first classification score and the second classification score to obtain a total classification score for the feather to be classified.

[0128] In some embodiments, the system may further include: a seventh module for creating the feather segmentation model based on feather images labeled with feather morphology tags and feather root morphology tags; and an eighth module for loading a pre-trained ResNet-152 model using the PyTorch deep learning framework and replacing the output layer of the ResNet-152 model with a multi-level classifier to obtain the feather classification model.

[0129] The content of the method embodiments of the present invention is applicable to the system embodiments. The specific functions implemented in the system embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0130] like Figure 9 As shown, another aspect of the present invention provides an electronic device 800, including a processor 810 and a memory 820;

[0131] Memory 820 is used to store programs;

[0132] The processor 810 executes programs as described above.

[0133] The content of the method embodiments of the present invention is applicable to the embodiments of the present electronic device. The specific functions implemented by the embodiments of the present electronic device are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0134] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the method described above.

[0135] The content of the method embodiments of the present invention is applicable to the computer-readable storage medium embodiments. The specific functions implemented by the computer-readable storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0136] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the aforementioned method.

[0137] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.

[0138] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0139] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0140] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution means, apparatus, or device (such as a computer-based device, a processor-including device, or other means that can fetch and execute instructions from, or in conjunction with, an instruction execution means, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution means, apparatus, or device.

[0141] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0142] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution device. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0143] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0144] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0145] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A deep learning-based feather shape classification method, characterized in that, include: Obtain the external image of the feathers to be classified; The shape image is preprocessed to obtain a feather image; The feather image is segmented using a feather segmentation model to obtain a feather shape image and a feather root shape image. The feather segmentation model is obtained by training the Unet semantic segmentation network on feather images labeled with feather morphology and feather root morphology. Using a feather classification model, the feather morphology images are classified into morphological levels to obtain a first classification score. The feather classification model is constructed based on a residual neural network. The output layer of the residual neural network is a multi-level classifier, and each level of the classifier corresponds to a different morphological level. The method further includes: A dataset was constructed by preprocessing feather root morphology images; Using the PyTorch deep learning framework, a pre-trained ResNet-152 model is loaded, and the output layer of the ResNet-152 model is replaced with a multi-level classifier. The model is then trained on the dataset to optimize the parameters of the classifier and obtain the feather classification model. The feather classification model is used to infer the defect level of the feather morphology image and obtain the defect level and the corresponding score. The feather root image is converted to the target HSV color space, and then the second classification score is obtained through color thresholding. The first classification score and the second classification score are weighted and calculated to obtain the total classification score of the feather to be classified.

2. The feather shape classification method based on deep learning according to claim 1, characterized in that, The preprocessing of the shape image includes at least one of the following: The shape image is subjected to image enhancement processing; The shape image is denoised. The shape image is cropped based on a preset size.

3. The feather shape classification method based on deep learning according to claim 1, characterized in that, The method further includes: The feather segmentation model is created based on feather images labeled with feather morphology tags and feather root morphology tags.

4. The feather shape classification method based on deep learning according to claim 3, characterized in that, The process of creating a feather segmentation model based on feather images labeled with feather morphology tags and feather root morphology tags includes: A training set is determined based on feather images labeled with feather morphology and feather root morphology; then, data augmentation is performed on the training set. The Unet semantic segmentation network is set up, and the Unet semantic segmentation network is trained for classification using the augmented training set. Based on the training results, the Unet semantic segmentation network is adjusted to obtain the feather segmentation model.

5. The deep learning-based feather shape classification method according to claim 4, characterized in that, The data augmentation of the training set includes at least one of the following: The images in the training set are randomly cropped to generate new training samples which are then added to the training set. The images in the training set are randomly flipped to generate new training samples which are then added to the training set. The images in the training set are randomly scaled to generate new training samples which are then added to the training set.

6. The feather shape classification method based on deep learning according to claim 1, characterized in that, The step of converting the feather root morphology image to the target HSV color space, and then obtaining the second classification score through color thresholding, includes: The feather shape image is converted to the target HSV color space to obtain the converted image; Based on a preset color threshold, pixel extraction is performed on the converted image to obtain a binary image; The binary image is filtered by area to obtain a filtered image; The selected images are divided into color levels by preset different color threshold levels to obtain the second category score.

7. A deep learning-based feather shape classification system, characterized in that, The system applied to the deep learning-based feather shape classification method of claim 1, the system comprising: The first module is used to acquire images of the shape of the feathers to be classified; The second module is used to preprocess the shape image to obtain a feather image; The third module is used to segment the feather image using a feather segmentation model to obtain a feather shape image and a feather root shape image. The feather segmentation model is obtained by training the Unet semantic segmentation network on feather images labeled with feather morphology and feather root morphology. The fourth module is used to classify the feather morphology image into morphological levels using a feather classification model to obtain a first classification score. The feather classification model is constructed based on a residual neural network. The output layer of the residual neural network is a multi-level classifier, and each level of the classifier corresponds to a different morphological level. The fifth module is used to convert the feather root morphology image to the target HSV color space, and then obtain the second classification score value through color thresholding. The sixth module is used to perform a weighted calculation on the first classification score and the second classification score to obtain the total classification score of the feathers to be classified.

8. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 6.

Images