Research On Complex Texture Image Defect Detection

Defect detection refers to locating the position of the defects in the product images and identifying the category of the corresponding position.It is a crucial process in the production process,which is conducive to controlling the quality of the product and reflecting the problems promptly.The defect detection of images with complex texture is the major problem in industrial visual detection,existing some problems such as having the complex background information in images,the feature informations are analogous to the pattern and the area of defects are so small that hard to detection.Therefore,it is significance for researching the algorithms for defect detection with complex texture images.In view of the problems existing in defect detection,based on the technology of deep-learning,this paper focuses on two key issues: the research of algorithms for defect detection and the lightweight of convolutional neural network.The research content mainly includes the following two points:(1)This paper proposes SA-RFB Net for detecting defects in complex texture images.Fisrtly,in view of the problem of having the complex background information in images and the feature information of defects are difficult to be extracted,this paper uses non-local module in order to capture long-range dependencies that could enhance the extraction capability.Secondly,in view of the problem of the feature informations are analogous to the pattern,this paper uses focal loss in order to put more attention on hard and misclassified examples and address the class imbalance during training a detector.In addition,in view of the problem of the area of defects are so small that hard to detection,SA-RFB Net uses multi-scale feature maps for detection,which improves the ability of detecting the small defects.(2)This paper proposes Light-SA-RFB Net to realize the compression and acceleration of SA-RFB Net.In view of the problem of the parameters and calculations of the SA-RFB Net could be optimized,this paper proposes the Ghost-RFB module and the Ghost-SA module to achieve the lightweight of SA-RFB Net.Light-SA-RFB Net is a lightweight network,which considers the efficiency while keeping the detection performance and could meet the requires of the actual industrial scenarios.In this paper,the ablation experiments and the comparison experiments of SA-RFB Net are carried out in the defect data set,and the ablation experiments and the comparison experiments of Light-SA-RFB Net are also carried out.The experimental results show that SA-RFB Net could accurately locate the position of the defects in the complex texture images and correctly identify the category of the corresponding position.Light-SA-RFB Net could further reduce the quantity of parameters and calculations while maintaining the detection performance,making defect detection more intelligent and convenient.