Research On Real-time Detection And Classification Of Metal Sheet Surface Defects Based On Machine Vision

Surface defect detection and classification is one of the most important research directions in the field of machine vision,and it acts a pivotal part in scenarios such as industrial quality inspection,chip manufacturing,intelligent transportation,medical diagnosis,etc.Surface defect detection is to use the feature information of the defect target in the image to determine whether there are defects on the surface of the object,and to further determine the position and size of the defects.Defect classification is to distinguish defect categories according to different defect characteristics reflected in the image.In practical applications,surface defect detection and classification algorithms are often required to have high detection accuracy,reasonable computational complexity,and small time overhead.In order to quickly and accurately determine various defect targets on the surface of an object,lots of algorithms have been proposed.However,traditional image processing algorithms are difficult to meet the latency and accuracy requirement of the real-time detection.In this thesis,micro metal sheets are used as the studied object,focusing on the key issues of real-time detection and classification of surface defects based on machine vision algorithms,and the following work is carried out:First,after summarizing and analyzing defect detection methods based on traditional image processing,a detection method based on connected domain analysis is proposed,which further classifies metal sheet surface defects according to feature types.This method makes full use of the gray value,color and shape of the defect target to perform morphological analysis,threshold segmentation,connected domain analysis and other operations on the metal sheet image to obtain the minimum circumscribed rectangle information of the defect area,thereby realizing the image segmentation of the defective area.Experimental results show that the proposed methods can achieve high accuracy for samples with continuous defect morphology characteristics and high contrast between defect targets and background regions.However,the computation intensity of the existing methods is large,the computation throughput and latency cannot meet the requirement of the real-time detection and needs to be further optimized.To overcome the long latency and low throughput of the connected domain analysis method,a novel surface defect classification method based on compact convolutional neural network(CNN)is proposed.The proposed method can automatically learn the deep feature information of the image,which greatly improves the ability to classify surface defects.The compact model proposed in this thesis contains 5 convolutional layers,3 pooling layers,rectified linear units,batch normalization layers,and Softmax function for the classification results.The surface defect classification method based on the compact CNN takes 7ms to classify a single image under the premise with the recognition accuracy rate of 96.85%.The prior connected domain analysis algorithm takes 59.55ms.Therefore,compact convolutional neural network based algorithm decreased the latency of defect detection and classification part within the image processing pipeline by 8.69x.For the entire detection pipeline of a single image,the latency decreased by 1.79x.Experiments have verified that this method can maintain a high recognition accuracy rate while meeting the classification latency requirements.It achieves the expected goal of an effective and rapid screening of metal sheet surface defects.Aiming at the real-time verification of the surface defect classification method of the compact CNN,this thesis designs and develops an online sorting of metal sheet surface defects software platform to realize the functions of online displaying of pictures,result and data statistics analyzing,and the compact convolution neural network computation.The neural network method is applied in the software platform of online chip sorting equipment,and the effectiveness and real-time performance of the compact CNN are verified.The software platform is mainly to obtain the metal sheet image through the image acquisition module,determine the defect type of the metal sheet surface according to the compact CNN model,and then feed the result back to the control system to realize automatic sorting.The software platform achieves a high degree of automation,meeting the real-time demand of 12-20 pieces per second.To sum up,in this thesis,a compact CNN based defect detection and classification algorithm is applied in the metal sheet surface defect sorting software platform.The experimental results show that the platform can realize real-time and effective screening of metal sheet surface defect,detection and it shows great practical value in a wide range of industrial screening applications.