Products' Surface Defects Recognition Method Based On Saliency Detection And OC-SVM

Surface defects detection is the prominent technique of quality assurance in the process of manufacturing.The crack,scratch and black spot defects can be detected by using the image processing,image acquisition,and so on.The coating product is researched in this article,and we aim to propose a relatively more intelligent method to detect the surface defects based on saliency detection and machine learning.Coating products are web-like structures spliced from steel wire through welding process and spray-coated with resin materials against rusting.Two types of defects are brought in the process of manufacturing including shape deformation defects and flaw defects,and the basic difference is that the size of the former is generally larger than the latter.The difficulty of surface defect detection for coating products is that their surfaces are 3D cylindrical structures,the surface brightness after imaging becomes heterogeneous and the size-ratio between flaw defects and the product reaching to 1/10000.Considering the fact that shape deformation defects are more dependent to the template image,so this work chooses global image registration for detecting these defects.Moreover to the flaw defects,these defects have smaller size,and its shape and position are random.Therefore this article proposes a saliency detection method based on self-subtraction algorithm to detect them.By this method,all the real-defects can be detected and meanwhile the result inevitably includes a part of pseudo-defects.Therefore in order to detect real-defects we need to classify defects.To improve the accuracy of the classification between real-defects and pseudo-defects,we need to extract feature vectors using feature description method before the classification,and its aim is to reduce the dimensionality of the source image.In this work we introduce three different feature description methods,which are respectively based on gray level co-occurrence matrix,Hu invariant moment,and gray level histogram.And they can respectively extract texture,shape and gray feature,and we conduct experiments to illustrate their detection effects.In order to choose an appropriate model,first we research defects detection applications based on SVM model.Second considering the fact that the defect-free images are easy to be collected and the defect images are difficult to be collected,therefore this article proposes to use OC-SVM as the defect recognition model.In the process of the model training,it only needs to learn defect-free images,which can overcome the disadvantages of the supervised learning method.To research and verify the performance and effectiveness of the algorithm,this article conducts experiments to test the effect of three different models including K-means,SVM and OC-SVM,and the experiment result shows that OC-SVM model is superior to others in terms of the overall accuracy.Besides this article uses experiments to find how the parameters of the OC-SVM model influence the detection result.Finally this article locally and globally tests the algorithm detection effect using a complete product image,and the result shows that our algorithm achieves a good result.