Image segmentation using improved JSEG

Abstract

Multi-class image semantic segmentation (MCISS) is one of the most crucial steps toward many applications such as image editing and content-based image retrieval. It's a very efficient method that include top down and bottom up approaches. In the top down approach model based segmentation is done. Semantic segmentation of image is one which groups the pixels together having common semantic meaning. This is done by applying semantic rules on the image pixels. Semantic texton forest (STF) is used for implementing this approach. In the bottom up approach using JSEG a region based segmentation is performed. To segment an input image, it heuristically groups the pixels in the input image according to their spatial adjacency, boundary continuity etc, and thus have no knowledge about the correspondence between pixels or regions to semantic categories, but will get more accurate boundaries than top down approach. But for some class of images JSEG showing reduced quality segmentation. To solve this FRACTAL JSEG method uses local fractal dimension of pixels as a homogeneity measure. This method showing improved result comparing to JSEG in boundary detection and hence segmentation. Another approach called I-FRAC also showing better results for some class of images where variation of colours is too low. Hence in this work an approach that uses both algorithms based on a selection criteria is proposed. This work is based on the assumption that by improving the bottom up approach using fractal dimension concept segmentation accuracy of MCISS can be improved. Here in the bottom up approach an improved version of JSEG is implemented to focus on how to find out a class specific value for region merging parameter that will increase the accuracy of segmentation.