Illumination invariant color image segmentation integrating visible color difference with edge information

Abstract

Image segmentation in variable illumination has great importance due to its wide range of appli- cations. Though there are many approaches for image segmentation, few of them works for images with variable illumination. Existing algorithms used for segmenting these images faces two major problems which are over-segmentation and failure to detect objects. This thesis is furnished with a novel algorithm called Illumination Invariant Color Image Seg- mentation Integrating Visible Color Difference with Edge Information. This algorithm manages varying illumination in images by incorporating the minimum color difference below which human being is un- able to detect the color discrepancy between two colors placed side by side. The work has two major contributions. The first one is to remove over-segmentation. The idea of minimum color difference is used to define homogeneous regions called visible color difference (VCD) regions which eventually minimizes the over-segmentation in the area of images with variable illumination. The second one is to identify the unidentified regions. Our method effectively exploits the edge information to identify the regions which other method fails to detect due to illumination variation in an image. Edge information ensures proper boundary of the newly generated segments as well. Thus the proposed segmentation algorithm substantially reduces the over-segmentation of images and also identifies the unidentified ob- jects that occur in many situation. We have developed the algorithm which uses no manual parameter provided during the execution time. Comparison with other algorithms has shown both quantitative and qualitative improvement of the proposed algorithm using Corel, google and Berkeley manual segmentation database. Based on F- measure the proposed algorithm has shown 11% better improvement over a recently developed variants of JSEG, called Fractal-JSEG.