Hybrid Region and Edge Based Unsupervised Color-Texture Segmentation for Natural Images

The paper proposes a generic color-texture feature integration framework. We propose two variants of edge based texture capturing method using filter banks of tensor products obtained from Orthogonal Polynomials (OP) OP3 of order 3 and OP5 of higher order 5 which are applied on Hybrid Color Space (HCS) for color texture feature integration. A region based unsupervised segmentation algorithm is applied on each of the variant’s Adaptive Feature Vector Representation (AFVR). The segmentation starts with classical K-means clustering in an iterative manner controlled by Kolmogrov-Smirnov (KS) test. A spatially constrained merge step acts as a post processing step to address over-segmentation. A segmentation tree is constructed in a hierarchical fashion generating metadata for cluster and regions separately. The algorithm is successfully tested quantitatively on 300 natural images from Berkley Standard Dataset using Probability Rand Index (PRI), Boundary Displacement Error (BDE), Variance of Information (VOI) and Global Consistency Error (GCE) which are widely used segmentation metrics found in literature. Experimental evidences are gathered showcasing the strength of color texture segmentation using OP3-HCS, OP5-HCS when compared with only color feature segmentation. A case study analysis indicates OP5 is biased towards over-segmentation as compared to OP3. Experimental results demonstrate the inherent simplicity and effectiveness of the proposed OP-HCS based hybrid color-texture image segmentation by achieving an average of 74 percent on PRI and at the same time having good balance on the rest of the three BDE, GCE and VOI measures. The results have been compared with other segmentation methods and found to be competitive.