Abstract
The paper presents two methods for similarity measurement of plastic textured images, and also for identification and localization of defective region in textured images. In order to have less sensitive statistic features regarding to rotated image, we introduce the notion of average co-occurrence matrix. With the purpose of algorithm validation, the images are divided into four equivalent regions. For the proper region similarity measurement, a decision theoretic method is used. In the fractal approach, we consider a new fractal dimension derived from box-counting algorithm, named effective fractal dimension, with an increasing efficiency for texture classification. Two experimental studies, one for statistical features and one for fractal type features, in a plastic simulated wood case, validate the algorithms. The algorithms are implemented in Visual C++ and Matlab. They allow the simultaneously display of both the investigated region, and the Euclidian distance between this and a reference image. The results confirm the fact that the distances between the regions without defect are relatively small, and the distance between a region with defect and a region without defect is relatively large. Also, the results show that features extracted from average co-occurrence matrix and the effective fractal dimension have a good discriminating power. Keywords: texture similarity, statistic features, fractal dimension, plastic simulated wood, defect detection