Small Sample Visual Defect Detection Based On DCGAN

As a big manufacturing country,Chinese booming manufacturing industry is closely related to the national economy.China has proposed the concept of intelligent manufacturing to meet the market demand for high-quality products because of the continuous development of artificial intelligence technology.The proposition of intelligent manufacturing can improve the traditional manufacturing mode and adopt more intelligent means to improve production efficiency and product quality.The surface quality is one of the most important factors to measure the quality of products.In order to reduce the economic losses and insecurity that may occur in defective products,many industries need comprehensive inspection of surface quality.Defect detection has always been an important part of the company's production of products,defect detection can enable enterprises to find defective products early,locate the cause,and avoid greater economic losses.The traditional defect detection methods are more manual,the detection process takes a long time,and the detection standards vary from person to person,so the more automated detection method has become the research focus of the enterprise.With the development of deep learning,the defect detection algorithm based on deep learning has emerged and achieved very good results,and it has been gradually applied in various industries.Based on the deep learning method,the appearance of the product is detected,the process is simplified,the time is short,and the accuracy is high.At the same time,more and more models for target detection and classification are proposed,which makes it possible to perform defect detection based on the deep learning method.At present,defect detection methods based on deep learning need to rely on a large number of defect samples.It is hard to achieve high precision defect detection effect with a limited number of samples.Aiming at this problem,this paper proposes a surface quality detection method for the small samples based on DCGAN(Deep Convolutional Generative Adversarial Network)sample generation.Firstly,a method of multiple types of defect patch generation based on DCGAN model is proposed,the method uses the original defect patch to train the DCGAN model to generate a number of new defect patch of the same type;Then,the method of sample set generation based on region growing and pre-location is proposed for the generation of defect patch data set,which is used to construct the defect image data set,effectively solve the problem of merging defect patch and original image fusion;Finally,Faster R-CNN is adopted to achieve the detection and classification of various types of surface defects.In this paper,three kinds of defects of precision workpieces,cracks,scratches and spots,are used to verify our method by experiments.Based on the original small samples of defect patches and the number of generated defect patches is 26700,the experiment results demonstrate that the accuracy of the classification of cracks,scratches and spots is more than 90% by the Faster R-CNN trained on generated defect patches.