A classification method of glass defect based on multiresolution and information fusion

The existence of defects is a key factor for quality degradation of float glass. This paper introduces an approach of glass defect identification based on multiresolution and information fusion analysis. With the help of 2D discrete wavelet transform, the subtracting defect image with a valid region is decomposed into approximated subimages and detailed subimages. The approximated subimages in three-level scales and the original defect image are chosen to proceed recognition with artificial neural network and fuzzy k-nearest neighbor. The decisive vectors from four classifiers are fused by an improved Dempster–Shafer (DS) evidence theory with head difference calibration-DS. Besides, a twice OTSU segmentation method as well as ten statistic features are interpreted for the preparation of defect recognition. The results of application indicate that the proposed algorithm can significantly increase the correct recognition rate of glass defects in five classes.