OBJECT DETECTION AND RECOGNITION BASED ON MULTISCALE DEFORMABLE TEMPLATE

A number of different algorithms have been described in the literature about template matching. Nevertheless, up to now no satisfactory solutions have been found to be both fast and immune to image deformation. A new approach based on a multiscale deformable template is presented. Compared with the basic active contour models, a shape restriction is added. Also, the multiscale property of wavelet transform is used to lead the matching process implemented from a coarse-to-fine scale. The initialization of the contour is done in a coarse scale, which can reduce the missing possibility and the compute complexity. The processing procedure includes three steps. First, multiscale edges of an image are detected based on wavelet transform; second, object contour is initiated in the input image by comparing Hausdorff distance in the coarsest scale; finally, the exact object contour is extracted via a coarse-to-fine scale implementation of the deformable template matching scheme. Experimental results with various kinds of image data have proved the algorithm to be efficient, precise and very immune to image deformation, noise and occlusion.