Object localization using deformable templates

TL;DR: A new algorithm is presented for localizing objects in images, using deformable templates, using a prototype template, which consists of shape-representative contours and edges, as well as a set of control points for image warping.

Abstract: A new algorithm is presented for localizing objects in images, using deformable templates. Prior knowledge of object shape is described by a prototype template, which consists of shape-representative contours and edges, as well as a set of control points for image warping. Computational efficiency is achieved using a multistage approach to find a match between the deformed template and objects in the image, by minimizing a cost function between the template and object boundary. The first stage of the algorithm reduces the physical search space size by determining the regions of interest using cross-correlation between the template and the edges of the image. A multiresolution paradigm is adopted for the second and third stages. In the second stage, an adapted hierarchical Chamfer matching scheme is used to find approximate matches between the template and the image using directional Edge Potential Fields (EPFs) of progressively higher resolutions. In the third stage, an innovative method using Particle Swarm Optimization is employed to find optimal control point placement at each resolution. A Local Weighted Mean (LWM) warp is also employed at this stage to facilitate a registration that iteratively deforms the template to fit these optimized points. The dimensionality of possible warp transformations is overcome by minimizing a cost function that penalizes extreme warps. The algorithm is succesfully applied to a number of images and the localization results are given, with each test set highlighting a different aspect of the algorithm.