Fabric Defect Detection by Singular Value Decomposition based Reduced Coefficient Fabric Space Optimized by Particle Swarm Technique and Implemented through OpenCL

To detect defects in woven fabric a reduced coefficient fabric space is constructed, for which the principal components representing row and column wise data distribution of the training fabric sub images are determined with the help of a novel singular value decomposition based method. The size of this reduced coefficient fabric space is suitably optimized by particle swarm optimization method, subject to the minimum detection error of fabric defects. For faster operation the generation of overlapped test fabric sub images, equal in size with training fabric sub images along with its projection on reduced coefficient fabric space is done by using parallel processing implemented through a novel algorithm in OpenCL. Finally detection of fabric defect is done by using support vector machine classifier. The method is tested on TILDA database for its validation. Significant performance enhancement is achieved as compared to 2 directional 2DPCA method both in terms of defect detection accuracy and reduction at computational cost.