Research On GAN-based Method For Augmentation Of Industrial Surface Defect Samples

In the field of industrial surface defect detection,it is usually necessary to collect a large number of samples and annotations for training deep learning model in subsequent detection phase.However,in actual application scenarios,due to the low probability of some types of defects occurring in the product,it is difficult to collect enough defect samples,and the effective detection effect of the deep learning model requires a large number of samples as support,which causes a huge challenge to improving the accuracy of surface defect detection models.To solve this problem,this thesis proposes an augmentation method for industrial surface defects,which uses the common industrial non-defect samples for data enhancement,and obtains the defect samples and corresponding segmentation annotations.The two aspects of research contents are as follows:(1)For the question of insufficient defect samples in industrial surface defect detection,this paper designs a surface defect image generation network Res-Pix2 Pix based on the Pix2 Pix structure,which uses a large number of easily obtained non-defect samples and a small number of defect samples and their annotations for training.The purpose of model training is to generate defects in the non-defect sample annotation area.The Res-Pix2 Pix model proposed in this thesis has made many improvements,the least square loss,the residual module and the label embedding module are added to improve the stability of training and the quality of generated defective images.Finally,through the defect generation experiment and the contrast experiment of the other enhancement methods,it is proved that the enhancement effect of ours is better than the Pix2 Pix model and geometric enhancement method.(2)Since Res-Pix2 Pix requires a certain amount of defect annotations as input when generating defect samples,but the probability of occurrence of defects is relatively low,so there are fewer defect annotations can be obtained.At this time,although the annotation of geometric transformation can be extended,the increase in annotation features is limited.To address this question,we present a VAE-based defect annotation enhancement network,which solves the problem of insufficient diversity of the defect annotation by training the original annotations.Finally,through the annotation generation experiment and the data enhancement experiment comparison with the original annotation,it is proved that the generated annotation has more advantages in the enhancement of industrial surface defects.Through the above two parts of research,the generation of industrial surface defects from non-defective samples to defect samples and their corresponding segmentation labels has been realized,which proves the effectiveness of our method in some application scenarios,and also provides a new idea for industrial surface defect enhancement.