DIGITAL IMAGE MATCHING METHOD USING NORMALIZED CROSS-CORRELATION (NCC)

Abstract

Digital image-matching techniques fall into three general categories: area-based, feature-based, and hybrid methods. In this experiment we use an area-based method that is Normalized Cross-Correlation (NCC) technique. In this approach, a statistical comparison is computed from digital numbers taken from same-size subarrays in the left and right images. The correlation coefficient can range from -1 to +1, with +1 indicating perfect correlation (exact match). a threshold value, such as 0.7, is chosen and if he correlation coefficient exceeds that value, the subarrays are assumed to match

The results present that using threshold=0.7, we can conclude that the similarity of two images using window size 3x3, 5x5, 7x7, 9x9 is 100%. But the similarity of two images using window size 11x11 is 70%. Then, using window size 3x3 and 5x5, we can achieve low RMSerror. It means that the different of position before and after NCC computation is not significant. However, using window size more than 8x8, RMSerror is so high more than 1 pixel.

The conclusions of this experiment are gaining larger window size, the correlation coefficient will be lower.  It means that the similarity is low and the images are not exactly match. Then, the large of window size give more number of grey value so it give effect in computation of average, then it will make the position of new pixel moved.  Getting larger window size so the movement will be so larger too, then the RMSerror will be high. And, the location of point will give contribution for the correlation coefficient. Because it will determine the varieties of pixel number