Best Paper Award 2013

Best Paper Award

Applications to Pattern Matching Using Spectral Theory and Its Performance Evaluation

Gou Koutaki , Keiichi Uchimura

[Trans. Inf. & Syst.（JPN Edition）, Vol. J96-D No.8, Aug. 2013]

Gou Koutaki

Keiichi Uchimura

　In the computer vision domain, pattern matching is one of the most popular approaches for detecting sought objects in images. Template and feature matching, such as image correlation and SIFT, are well-known forms of pattern matching in this domain. In using such methods, we need to prepare a number of images with various scales, sizes, and rotations in order to achieve invariance against such transformation. However, this entails high computation costs and can produce deterioration in matching accuracy.
　As a means of overcoming these difficulties, this paper presents an infinite-dimensional principal component analysis for pattern matching. In practice, the eigenspaces of the Gaussian-scale and Scale-Normalized LoG spaces defined by infinite different scaling images are derived from spectral theory. As applications, this paper introduces blurred image generation and SIFT-based feature matching. The results show that the proposed method is effective for those applications. Thus, it can be concluded that this paper is worthy of recognition with the Best Paper Award.