Feature Extraction and Recognition Method of Surface Defects Based on Independent Component Analysis

To extract features of surface defects,linear transformations are often performed on images.However,general wavelet and Gabor transformations are pre-defined and unchangeable,and their basis functions can not adapt to the characteristics of defect images.In this paper,a feature extraction method based on independent component analysis（ICA） and topographic independent component analysis（TICA） is proposed and applied to automatic recognition of cold rolled steel strip surface defects.Firstly the basis functions and filters which adapt to the characteristics of defect images are estimated adaptively using ICA and TICA from the defect library.Then the defect images are filtered to produce feature vectors.Finally the samples are classified by the support vector machine.The method is established on the basis of learning the defect library,being capable of extracting adaptively salient features of defect images.It has low computational complexity and high computational parallelism.Experimental results demonstrate that the method proposed has very high recognition rates for shape defects,texture defects and other defects.The total recognition rate can be as high as 95.52%.