Object recognition based on fractal neighbor distance

Abstract

We have investigated a new method of object recognition based on fractal image coding. Fractal image coding can approximate any given image by capturing intrinsic self-similarities within the image. A database of fractal codes of training images was created, and for each input image the input–output characteristics of each fractal code was measured using the Euclidean norm. The input image was assigned to the class of fractal codes that minimized this norm. The contractivity factor of the fractal code and the encoding scheme used was shown to affect recognition rates. This method was applied to face recognition. The performance of several variants of this algorithm was compared to other approaches to face recognition including eigenfaces, and the nearest neighbor classifier. The best variant of this new method achieved an average error rate of 1.75% on the publicly available Olivetti Research Laboratory face database with an average classification time of 3.2 s .