Research On Few-shot Image Recognition Method For Industrial Product Surface Defects

With the rise and development of artificial intelligence technology,machine learning methods are also widely used in practical industrial production scenarios.The efficiency of industrial product surface defect recognition based on machine learning methods is much greater than that of traditional manual or equipment recognition methods.With the continuous progress of related research,the performance of machine learning methods in industrial surface defect recognition application scenarios has been growing,however,in practical applications face the problem of few-shot for image recognition,i.e.the number of defect samples for machine learning method training is too small,in extreme cases only few samples per class.To address this problem,this thesis introduces few-shot machine learning methods and cross-domain ideas,and carries out research on image classification and image generation for industrial product surface defects respectively,in order to solve the problem of industrial product surface defect recognition in practical applications from different perspectives.The thesis is based on the research cooperation project "Research and development of edge intelligence technology and system equipment for smart factories",which aims to solve the technical problems faced by the implementation of artificial intelligence for smart factories from the perspective of practical applications.The main research contents of this thesis are as follows.1)A review of the current state of research on the few-shot problem in industrial product surface defect image recognition.The definition and application scenarios of the few-shot problem are firstly sorted out,and then the research background and research directions of industrial product surface defect recognition are summarized.The research status of image classification methods and image augmentation methods in the few-shot problem of image recognition are reviewed respectively,the data set requirements and training methods in the existing research are given,and the shortcomings and limitations of the existing research in generating images and other aspects are analyzed and discussed,which provide ideas for the subsequent research of this paper.2)A cross-domain few-shot image classification method based on representative multi-domain feature selection is proposed for the problem of few-shot image classification for industrial product surface defect recognition.The proposed method combines domain generalization with few-shot machine learning methods based on the idea of feature selection to cope with the few-shot image classification problem.Specifically,a domain generalization method is introduced to optimize the data preprocessing process in order to better extract domain-general features;a local attention-based feature-wise linear modulation module is designed to be inserted into the feature extractor to achieve domain-specific feature extraction;and a multi-loss dynamic weighting method is designed to optimize the feature selection process in conjunction with a prototype network approach.Experimental validation is conducted on a publicly available baseline dataset and an actual industrial product defect dataset.The results show that the proposed method has superior performance in terms of classification accuracy compared with existing methods in fewshot scenarios,with an improvement of about 3%in classification accuracy.3)To address the problem that machine learning-based methods for industrial product surface defect recognition require a large number of image samples,an image generation method based on generative adversarial networks is proposed.The proposed method is based on crossdomain and meta-learning ideas enabling the proposed method to cope with the problem of few-shot in real industrial scenarios by augmenting the number of surface defect image samples.In the pre-training phase,a style block-based backbone network is designed and fully pre-trained on the source domain to extract domain-general feature information.In the model training phase,the model is trained using the episode training method in meta-learning,where each episode is divided into an inner optimization and an outer optimization,so that the model can be adequately trained when the number of image samples is sparse,and eventually generates a large number of image samples required by the machine learning method.Experimental validation on a real-world dataset of industrial product surface defects shows that the proposed method has performance advantages over existing methods in terms of the quality and quantity of images generated.