Object localization using deformable templates

Abstract

Object localization refers to the detection, matching and segmentation of objects in images. The localization model presented in this paper relies on deformable templates to match objects based on shape alone. The shape structure is captured by a prototype template consisting of hand-drawn edges and contours representing the object to be localized. A multistage, multiresolution algorithm is utilized to reduce the computational intensity of the search. The first stage reduces the physical search space dimensions using correlation to determine the regions of interest where a match it likely to occur. The second stage finds approximate matches between the template and target image at progressively finer resolutions, by attracting the template to salient image features using Edge Potential Fields. The third stage entails the use of evolutionary optimization to determine control point placement for a Local Weighted Mean warp, which deforms the template to fit the object boundaries. Results are presented for a number of applications, showing the successful localization of various objects. The algorithm’s invariance to rotation, scale, translation and moderate shape variation of the target objects is clearly illustrated.