New Approach of Texture Classification and Retrieval using Scale Invariant Feature Transform

Van Tuan Do
Dept. of Computer Engineering and IT, University of Ulsan

*Anastasiya Yun
Dept. of Computer Engineering and IT, University of Ulsan

*Jong-Soo Lee
Dept. of Computer Engineering and IT, University of Ulsan

*Ui-Pil Chong
Dept. of Computer Engineering and IT, University of Ulsan

Abstract
A texture image can be viewed as a collection of significant parts, which are presented based on local image features. In this paper, we built local image features (also known as descriptors) using the Scale Invariant Feature Transform (hereinafter SIFT) algorithm from the training set of texture images. In general, a texture can be characterized through textons, which are formed by clustering those local image features. To establish the texton dictionary, an adaptive mean shift clustering algorithm is run on a randomly selected subset of all extracted descriptors. Once the texton dictionary is built, the texton histograms can be made for each image. The texton histograms are used to measure a similarity between images. Moreover, we implemented the SIFT algorithm with various parameters (number of scale space, number of different of Gaussian space and the masks of Gaussian) to have descriptors be suitable with texture characteristics.