Film product few-sample defect classification method based on deep learning
Through deep learning, the small sample learning technology combined with meta-learning and data enhancement, the problem of limited sample number in membrane product defect detection is solved, efficient and accurate defect classification is achieved, detection accuracy and efficiency are improved, cost is reduced, and enhanced The applicability of the model.
Patent Information
- Application Number
- CN202411622950.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-14
- Publication Date
- 2025-01-10
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Film products are prone to defects during production. The existing detection methods rely on manual or simple visual image processing, which are inefficient and have limited accuracy. Deep learning models require a large number of labeling samples, making it difficult to obtain defect samples, which limits their application.
Using a low-sample learning technology based on deep learning, features are extracted through convolutional neural networks, combined with meta-learning framework and data augmentation method, transfer learning and Voting model fusion are used to achieve accurate classification of membrane product defects.
It realizes efficient and accurate defect classification under a small number of samples, improves detection accuracy and efficiency, reduces labor costs, enhances the generalization and scope of application of the model, and is suitable for different types of defect detection tasks.
Smart Images
Figure CN119295837A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image processing and machine learning, and specifically to a method for classifying defects of a small number of membrane products based on deep learning. Background Art
[0002] With the continuous improvement of industrial automation, membrane products are increasingly used in water treatment, energy, medicine and other fields. However, membrane products are prone to defects during the production process, which can seriously affect the performance of the product. Defect detection methods rely on manual or simple visual image processing technology, which is not only inefficient but also has limited accuracy. In recent years, deep learning technology has made significant progress in the field of image recognition, providing new solutions for the automated detection of membrane product defects. However, deep learning models usually require a large number of labeled samples for training, and in practical applications, it is often difficult to obtain a large number of defect samples, which limits the application of deep learning technology in membrane product defect detection. Summary of the invention
[0003] The purpose of the present invention is to provide a membrane product few-sample defect classification method based on deep learning, aiming to solve the problem of limited number of samples in membrane product defect detection. The method utilizes few-sample learning technology to achieve accurate classification of membrane product defects by effectively learning a small number of labeled samples.
[0004] To achieve the above object, the present invention provides the following technical solutions: A method for classifying defects of membrane products with a small number of samples based on deep learning comprises the following steps: S1: Data collection and preparation: Collect image samples of film products with and without defects, and perform data cleaning and preprocessing, including image enhancement, denoising, and size normalization; S2: Feature extraction and representation: Using convolutional neural networks in deep learning, we extract and represent features of preprocessed images. These features can capture and express local and global information in the image. S3: Few-shot learning strategy: Introducing a meta-learning framework to train the model on multiple but different related tasks so that the model can learn how to learn quickly from small samples; S4: Defect classification: The image of the membrane product to be classified is input into the trained model, and the classification result is obtained through forward propagation calculation to determine whether it has defects and the type of defects.
[0005] Preferably, digital image processing methods and generative adversarial networks are used in S1 to perform data enhancement to increase the number of samples and improve the generalization ability of the model.
[0006] Preferably, in S2, a feature extractor of a migration deep learning model is used to input the enhanced data into the model for feature extraction.
[0007] Preferably, in S3, a new classification structure is trained to obtain a set of migration deep learning models, and a voting model fusion method is used to perform model fusion, and voting is performed in accordance with the principle of minority obeys majority to obtain a defect recognition result.
[0008] Preferably, the S4 utilizes the deep learning model DAGAN for data enhancement, which only requires a small number of samples to be used in the deep learning recognition model and can have a higher accuracy.
[0009] In summary, due to the adoption of the above technology, the beneficial effects of the present invention are: The core of the method of the present invention is to utilize the feature extraction ability of the deep learning model and the data enhancement technology to achieve efficient defect classification through a small amount of samples. In specific implementation, the feature map can be embedded in the encoder network structure, and the model's ability to characterize samples and the accuracy of pseudo-label annotation can be enhanced through the pseudo-label mechanism and the loss function of deep one-class classification anomaly detection.
[0010] 2. Efficient, accurate and customizable: Using deep learning technology, it can provide higher detection accuracy and faster speed than traditional methods, thereby improving detection efficiency. Flexible data pre-processing and post-processing steps make it easy to adapt to different types of defect detection tasks and have strong adaptability.
[0011] 3. Reduce labor costs and improve identification efficiency: Through automated testing, labor costs can be reduced, production efficiency can be improved, deployment is simple, repeatability is high, and defects can be effectively identified at low cost.
[0012] 4. Good model generalization: Through transfer learning, the model can achieve high generalization under the condition of few samples, thereby improving the applicability of the model; Good defect segmentation effect: By introducing unsupervised learning methods, only a small number of positive samples are needed for training to achieve pixel-level defect segmentation and improve defect segmentation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention, so that other features, purposes and advantages of the present invention become more obvious. The accompanying drawings of the exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings: Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0014] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the scope of protection of the present invention.
[0015] In the description of the present invention, it is necessary to understand that the terms indicating orientation or positional relationship are based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.
[0016] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances of the specification.
[0017] The present invention provides Figure 1 A method for classifying defects of membrane products with a small number of samples based on deep learning is shown, comprising the following steps: S1: Data collection and preparation: Collect images of defective and non-defective membrane products, and perform data cleaning and preprocessing, including image enhancement, denoising, and size standardization. Use digital image processing methods and generative adversarial networks to perform data enhancement to increase the number of samples and improve the generalization ability of the model. S2: Feature extraction and representation: Use convolutional neural networks in deep learning to extract and represent features of the preprocessed images. These features can capture and express local and global information in the image. Use the feature extractor of the transfer deep learning model to input the enhanced data into the model for feature extraction. S3: Few-sample learning strategy: Introduce a meta-learning framework to train the model on multiple but different related tasks, so that the model can learn how to quickly learn from small samples. By training a new classification structure, a set of transfer deep learning models is obtained. The voting model fusion method is used to fuse the models, and the majority principle is followed to vote to obtain defect recognition results. S4: Defect classification: Input the membrane product image to be classified into the trained model, and obtain the classification result through forward propagation calculation to determine whether it has defects and the type of defects. Use the deep learning model DAGAN for data enhancement. Only a small number of samples are needed to use it in the deep learning recognition model, and it can have a higher accuracy.
[0018] The core of this method is to use the feature extraction ability of the deep learning model and the data enhancement technology to achieve efficient defect classification with a small number of samples. In specific implementation, the feature map can be embedded in the encoder network structure, and the pseudo-label mechanism and the loss function of deep one-class classification anomaly detection can be used to enhance the model's ability to represent samples and the accuracy of pseudo-label annotation.
[0019] Specific application scenarios and effects: In the defect detection of membrane products, this method can effectively deal with the problems of limited sample size and high data annotation cost. By constructing a defect detection network model, using feature map embedding encoder network structure, combining pseudo label mechanism and balanced cross entropy function for training, the accuracy of defect detection can be significantly improved and the computational complexity can be reduced.
[0020] Future development trends: With the continuous development of deep learning technology, the few-sample defect classification method based on deep learning will be applied in more fields. In the future, with the continuous optimization of the algorithm and the improvement of hardware performance, this method will be more efficient and accurate, and suitable for defect detection tasks in more complex scenarios. Embodiment 1
[0021] Step 1: Capture and process membrane product images: Use an industrial camera to capture membrane product images, and adjust the height and focal length of the industrial camera to capture defective images of membrane products in the corresponding state. Reconstruct the image data of membrane products, including but not limited to steps such as denoising, grayscale, and standardization, in order to improve the efficiency and accuracy of subsequent model training. Enhance small samples through data enhancement techniques (such as rotation, scaling, rotation, etc.) to enhance the generalization ability of the model.
[0022] Step 2, feature extraction: A deep neural network (CNN) is used as a feature extractor to extract fine features of defects from the repaired image. Considering the characteristics of few-sample learning, a lightweight network structure suitable for few samples is selected to reduce computing resource consumption and increase processing speed.
[0023] Step 3: Few-sample learning strategy: Introduce the meta-learning framework, and train the model on multiple related but different tasks to enable the model to learn how to quickly learn new tasks from small samples. Use the model-related meta-learning algorithm (MAML) method to effectively adapt quickly from small samples and improve classification performance.
[0024] Step 4, defect classification: Based on data extraction and meta-learning, a classifier is constructed to classify the defects of membrane products. New membrane product images are trained for defect classification to achieve fast and accurate defect detection.
[0025] This approach has multiple benefits: 1. Improve detection accuracy: Using deep learning technology, this method can provide higher detection accuracy than traditional methods and effectively identify defects on the surface of membrane products; 2. Enhance model generalization: Through transfer learning and other technologies, this method can improve the generalization of the model under the condition of few samples and solve the problem of insufficient samples; 3. Reduce costs: This method can reduce labor costs, improve recognition efficiency, and reduce enterprise development costs, because deep learning models are customizable and some are open source and free; 4. Flexible adaptation: This method has flexible data preprocessing and post-processing steps, which is convenient for adapting to different types of defect detection tasks and meeting the diverse production needs of enterprises.
[0026] In summary, the membrane product few-sample defect classification method based on deep learning has significant advantages in improving detection accuracy, enhancing model generalization, reducing costs and flexible adaptation. Embodiment 2
[0027] 1. Data Collection and Preprocessing (1) Data collection: Collect images of film products from the production line containing various defect types.
[0028] Due to the few-shot case, the number of images for each defect type may be very limited (e.g., dozens or less).
[0029] (2) Data preprocessing: The collected images are cropped, rotated, scaled, and other transformations are performed to increase the diversity of the data.
[0030] Use image enhancement techniques such as contrast adjustment, noise addition, etc. to further enrich the dataset.
[0031] Convert the image to the input format required by the deep learning model (such as RGB format, specific size, etc.).
[0032] 2. Deep learning model selection and construction (1) Model selection: Choose a deep learning model suitable for few-shot learning, such as a transfer learning model (using a pre-trained convolutional neural network, such as VGG, ResNet, etc., and fine-tuning it on its basis) or a meta-learning model (such as MAML, Meta-SGD, etc.).
[0033] (2) Model construction: According to the characteristics of membrane product defects, adjust the model parameters such as the number of layers, convolution kernel size, and activation function.
[0034] Set a classifier at the output layer of the model that matches the number of defect types.
[0035] 3. Model training and optimization (1) Training strategy: Use data augmented images for training to improve the generalization ability of the model.
[0036] A cross-validation strategy was used to divide the limited dataset into a training set and a validation set to evaluate the performance of the model.
[0037] (2) Optimization method: Choose an appropriate optimizer (such as Adam, SGD, etc.) and learning rate scheduling strategy.
[0038] Use regularization techniques (such as L1, L2 regularization) and Dropout strategies to prevent the model from overfitting.
[0039] Monitor the loss function and accuracy during training and adjust model parameters in a timely manner.
[0040] 4. Model Evaluation and Testing (1) Evaluation indicators Choose appropriate evaluation metrics such as precision, recall, F1 score, etc. to comprehensively evaluate the performance of the model.
[0041] Use the confusion matrix to visualize the classification effect of the model and analyze the misclassification of various defects.
[0042] (2) Testing and verification Verify the performance of the model on an independent test set to ensure that the model has good generalization ability.
[0043] Compare the performance differences between different models or different training strategies and select the optimal model.
[0044] 5. Model deployment and application (1) Deployment strategy Deploy the trained model to the production line to achieve real-time membrane product defect classification.
[0045] Design a user-friendly interface to facilitate operators to view classification results and perform subsequent processing.
[0046] (2) Application scenarios Applied to the quality inspection process on the membrane product production line to improve inspection efficiency and accuracy.
[0047] Provide data support for product development and improvement, and help identify common defect types and causes.
[0048] 6. Continuous improvement and updating New defect image data is collected regularly to update and retrain the model.
[0049] Pay attention to new technologies and methods in the field of deep learning, and promptly apply advanced algorithms to membrane product defect classification.
[0050] Through the steps of the above-mentioned embodiment, a method for classifying defects of membrane products with a small number of samples based on deep learning can be constructed. This method can effectively classify defects on membrane products with a limited number of samples, thereby improving the quality inspection efficiency and accuracy of the production line.
Claims
1. A method for classifying defects of membrane products with a small number of samples based on deep learning, characterized in that: The following steps are involved: S1: Data collection and preparation: Collect image samples of film products with and without defects, and perform data cleaning and preprocessing, including image enhancement, denoising, and size normalization; S2: Feature extraction and representation: Using convolutional neural networks in deep learning, we extract and represent features of preprocessed images. These features can capture and express local and global information in the image. S3: Few-shot learning strategy: Introducing a meta-learning framework to train the model on multiple but different related tasks so that the model can learn how to learn quickly from small samples; S4: Defect classification: The image of the membrane product to be classified is input into the trained model, and the classification result is obtained through forward propagation calculation to determine whether it has defects and the type of defects.
2. The method for classifying defects of membrane products with a small number of samples based on deep learning according to claim 1, characterized in that: In S1, digital image processing methods and generative adversarial networks are used to perform data enhancement to increase the number of samples and improve the generalization ability of the model.
3. The method for classifying defects of membrane products with a small number of samples based on deep learning according to claim 1, characterized in that: In S2, a feature extractor of a transferred deep learning model is used to input the enhanced data into the model for feature extraction.
4. The method for classifying defects of membrane products with a small number of samples based on deep learning according to claim 1, characterized in that: In S3, a new classification structure is trained to obtain a set of migration deep learning models, and the model fusion is performed using the Voting model fusion method. Voting is performed in accordance with the principle of minority obeys majority to obtain the defect recognition result.
5. The method for classifying defects of membrane products with a small number of samples based on deep learning according to claim 1, characterized in that: The S4 uses the deep learning model DAGAN for data enhancement, and only requires a small number of samples to use it in the deep learning recognition model, and can have a higher accuracy.