Research On Industrial Defect Detection Based On Deep Learning

As the screen quality of electronic products is a crucial concern of consumers in recent years,the defect detection,including screen background and defect characteristics,is necessary during industrial production.The traditional manual and machine vision detection might lead to lots of missed and false detection,for their shortcomings of low accuracy and efficiency.In contrast,the deep learning technology could discover the rules behind complex things,which excels in feature extraction and expression,as well as strong ability.Based on the deep learning technology,the present research aims to improve the traditional screen defect detection,by the algorithm of screen background and defect.The main contributions of the current research are listed below.Firstly,aiming at the problem that some defects and low background contrast cannot be detected,due to the complex and changeable industrial screen background,a screen background suppression algorithm,which is based on improved homomorphic filtering,is proposed in order to compress the changing range of screen background pixels,and to improve the defect and background contrast.Secondly,aiming at the weak generalization in different production lines,the numeral imbalance between good and defective samples in industrial scenes,as well as the failure of deep learning detection,an unsupervised screen background generation model is proposed,in order to generate the screen background template,as well as to improve the separation accuracy of the background and defects.Thirdly,aiming at the vague defect features which is difficult to be separated from the background,a deep residual reconstruction network model is proposed,in order to improve the feature connection structure of the residual block,to strengthen the feature extraction ability of the model under the condition of simplifying network parameters and reducing resource consumption,and to enhance the defect features.The fine qualities of the above scheme,including the highlighted defect features,the improved detection accuracy,as well as the reduced missing and false detection of defects,is proved in the present research.