Image Categorization and Semantic Segmentation using
Scale-Optimized Textons

Yousun Kang^1* and Akihiro Sugimoto²

¹Tokyo Polytechnic University

Atsugi 243-0297, Japan
yskang@t-kougei.ac.jp

²National Institute of Informatics
Tokyo 101-8430, Japan
sugimoto@nii.ac.jp

Abstract

In computer vision research, a texton is a representative dense visual word for the bag-of-keypoints method.
It has proven its effectiveness in categorizing materials and in generic object classes. Despite its success and popularity, no report describes a study that has tackled the problem of its scale optimization for given image data and associated object categories. We propose scale-optimized textons to learn the best scale for each object in a scene. We incorporate them into image categorization and semantic segmentation. Our textonization module produces a scale-optimized codebook of visual words. We approach the scale-optimization problem of textons using the scene-context scale in each image, which is the effective scale of local context to classify an image pixel in a scene. We perform the textonization process using a randomized decision forest, which is a powerful tool with high computational efficiency in vision applications. Results of our experiments using MSRC and VOC 2007 segmentation datasets demonstrate that our scale-optimized textons improve image categorization and segmentation performance.

Keywords: scale-optimized textons; image categorization; semantic segmentation; decision forest.

*Corresponding Author:
1583 Iiyama, Atsugi, Kanagawa 243-0297 Japan, Tel: +81-46-242-9524
Web: http://researchmap.jp/yskang/?lang=english

IT Convergence Practice (INPRA), Vol. 2, No. 1, pp. 2-14, March 2014 [pdf]