Research On Gan-based Defect Sample Generation And Defect Detection

When using deep learning method to complete the industrial defect detection task,there is a problem that the training data of defect samples is insufficient.To solve this problem,data augmentation for surface defect detection is an important research.At present,the research on data augmentation based on GAN has a good effect.Through adversarial training,the network can effectively learn the feature information of defect samples,and the generated results are closer to the real defect samples.However,there are also some problems such as poor background area generation and insufficient diversity of defect samples when using GAN for data augmentation.Aiming at the problems in the research of data augmentation for surface defect detection,based on the latest research of GAN,the thesis carries out the research on surface defect gnenration.The thesis proposes a GAN-based defect sample generation algorithm through the research on GAN and defect sample generation.In the proposed algorithm,the full convolution generator with U-shaped structure is designed,and the spatial adaptive normalization structure is introduced to control the generated defect shape by mask annotatio;multi-layer convolution discriminator is designed to extract the adversarial feature information between real samples and generated samples;the design of adversarial training loss completes the generator training.The thesis proposes a GAN-based surface defect detection algorithm through the research of defect detection network training by introducing defect free samples using GAN.In algorithm,the U-shaped structure defect editing network and multi-layer convolution discriminator are designed to complete the defect editing function from defect samples to defect free samples,and introduce defect free samples into network training;the loss of training is designed to complete defect editing network training;the residual binary module is designed to binarize the residual image of defect editing,and get the final semantic result.In the thesis,the GAN-based defect sample generation algorithm is tested on the LCD screen defect dataset and DAGM dataset.Through experiments,it is proved that the segmentation network has better segmentation results than that without data augmentation.Then,the GAN-based defect detection algorithm is tested on DAGM dataset and magnetic tile surface defect dataset.Through experiments,the m Io U of the proposed algorithm is 2.4% higher than that of Unet,and its segmentation stability for different types of defects is higher.