| Today,with the rapid development of science and technology in the world,China,as a major manufacturing country,strives to achieve high-quality development.And in the actual production process of enterprises,it is quite important to do a good job of identifying the surface defects of products.The use of manual identification,especially under the working conditions of special environments,can lead to inefficient identification,cause certain harm to the body of workers,is not conducive to product quality control,and does not meet the background of the era of high-quality development,so the defect identification based on machine vision is particularly important.In this paper,a defect recognition method based on width learning system and convolutional neural network is proposed.The details of the research are as follows.(1)First,the advantages and disadvantages of current surface defect recognition algorithms and their research value are verified,and the problems that occur are analysed.(2)Then,in response to the problem that deep learning takes a lot of time to extract high-level image features through convolutional operations,this paper proposes a defect recognition method based on a bread learning system for strip steel surface defects.Using the flexibility of the bread learning system,we can extract the texture features of the image from the grey scale co-occurrence matrix,and subject these features to PCA dimensionality reduction,and finally apply the fuzzy bread learning system to defect recognition to achieve more accurate classification.This method effectively improves the strip defect recognition time while maintaining a high recognition accuracy.(3)Finally,to address the problem that the features extracted by deep learning convolutional neural networks are local features and lack of wholeness,this paper proposes a method to fuse the local feature information extracted by convolutional neural networks with the global feature information extracted by visual transformers(Vi T).Using Rep VGG as the backbone network,a network structure is designed that enables the full interactive fusion of the two feature information.The model was validated and analysed using a strip defect dataset,and the model structure was able to achieve 100% recognition accuracy on85% of the training set and 98.5% accuracy on 70% of the training set,improving the recognition accuracy by 5-7 percentage points over the original Rep VGG network. |
