Reusing the Otsu threshold beyond segmentation

Abstract

The Otsu thresholding is a classical binarization method that partitions graylevel images according to a within-class variance minimization principle. The Otsu method is a particular case of the general Lloyd-Max optimal quantization. We propose the alternative use of Otsu/Lloyd thresholds, computed locally, as local features that describe the image content. This description can be used directly within a content-based image retrieval framework, or it can be reused in the definition of new Local Binary Pattern variants. Texture retrieval experiments show that the proposed approaches lead to performance improvement under specific constraints.