Research On Surface Defect Detection Algorithm Based On Deep Learning

Surface defect detection is vital in product quality control.As an efficient and reliable solution,automated surface defect detection technology based on machine vision has been extensively studied and gradually applied to production.In recent years,deep learning has made breakthroughs in the field of machine vision.In order to improve the accuracy of defect detection,introducing deep learning into surface defect detection has been one of the research hotspots in the field of machine vision.This paper studies the general surface defect detection method based on deep learning.The research focuses on two types of problems in surface defect detection,namely,the defect identification problem of discriminating defects and the defect segmentation problem of quantitative analysis of defects.Algorithm evaluation is performed on the self-made button defect dataset and three other public datasets.Aiming at defect identification problem,this paper studies a defect recognition algorithm based on fully convolutional neural network and block detection.This algorithm introduces the strategy of block detection and fully convolutional neural network into ResNet,and overcomes the lack of receptive field in traditional block detection while realizing the local discriminating mechanism.Compared with the original ResNet image classification algorithm,this algorithm has stronger generalization ability and ability to detect small defects.For defect identification tasks for button,welds and silicon steel strip,the comprehensive accuracy on multiple datasets is increased by 1.13%~8.27%.Aiming at the defect segmentation problem,a defect segmentation algorithm based on improved deeplab v3+ is proposed.According to the characteristics of the defect segmentation task,the improved algrithm emphasizes the use of low-level local features.The improved algorithm adds a branch network based on the deeplab v3+ network structure,and improves the backbone network and the decoding network.To implicit extend training samples and strengthen the generalization ability of the network,the training strategy of partial image input is proposed.Compared with the original deeplab v3+ algorithm,F1 score is increased by 2.65%~3.35% on defect segmentation tasks for button,road,welds and silicon steel strip.Aiming at the problem of insufficient defect training samples in the actual defect detection task,a defect simulation algorithm based on Generative Adversarial Nets is proposed.This paper proposes the simulation network with an encoding-decoding structure and the regional training strategy for local image generation.With sufficient adversarial training,the simulation network can generate simulation defects of the specified type and shape based on the positive samples.The experimental results on multiple datasets show that combined with the defect simulation algorithm,the requirement for the number of defect training samples is reduced for deep learning based defect detection model.