부산대학교 도서관

In this paper correlation-based matching solutions of images for the case of the validation of the eurocoins with relatively high-speed motion have been developed and evaluated. Image processing (combined with other — electromagnetic, optical and acoustical sensorics) is an efficient method for coin recognition and validation. From the image processing viewpoint — an universal method is finding and checking the coin image by cross-correlating it with expected (reference) image(s). Correlation is widely used as an effective similarity measure in matching tasks. However, traditional correlation based matching methods are limited by various ways. Main challenge here (additionally to "pre-conditioning" of compared images, of course) is the significant reduction of the computations, needed for finding cross-correlation values over 2-D space and for all possible rotation values. In current work usage of variable-rate subsampled (by pixel blocks) reference images has been proposed and evaluated. A special case has been considered, where most of the image blocks are sampled with zero (none) or one sample-values, while a few blocks (some to some tens) of the image has been sampled at the full accuracy (eg 240−240 pixels, in the used examples). The criteria, used for selection of the “specific” (unique) blocks of the reference image has been the minimum value of the maximum cross-correlation of an block of samples (eg 20−20 or 30−30 pixels) against any other (shifted or rotated) block of the same coin image. So, blocks with relatively high value of “secondary “correlation peaks at shifting and rotating are not considered as “good or unique” ones. So, sub-sampled reference images for various eurocoins has been proposed, and corresponding algorithms has been evaluated. Alternatively, using of set of feature-points of the images, as reference for cross-correlation, has been evaluated, for various eurocoins (with corresponding determination of reasonable feature points, for these coins). Alternatively, using local maximum and minimum difference based interesting pixel blocks as reference blocks for cross-correlation has been discussed. These methods have been tested on real eurocoins, the results are presented at the end of paper.