Color image segmentation via fuzzy feature tuning and feature adjustment

This paper presents a novel approach that incorporates watershed analysis, fuzzy feature tuning (FFT), and feature adjustment to perform color image segmentation. It is well known that human eyes do not respond to intensity variation in linear fashion (Gonzalez and Woods, 2002). Furthermore, we have observed that most existing segmentation algorithms treat color features independently, but the results likely contain false counters. Compared to gray image, color image segmentation is more difficult because the inherent multi-features not only contain nonlinear relations individually but also comprise inter-feature dependency between R, G, and B (or Y, C_b, C_r). Many existing segmentation algorithms ignored these two important facts and simply combine the three separate feature planes; making false contours inevitable. This work studies nonlinear dependency of features and exploits inter-feature and intra-feature information. The results are shown useful in adjusting input features while merging small regions that encompass false contours. The proposed approach mainly consists of two parts. First, feature planes of color images are individually segmented by watershed analysis, due to its easiness in implementation and capability of generating a complete closed contour. We show that the well-known over-segmentation problem (i.e., numerous initial regions) from using watershed analysis can be easily solved by the FFT algorithm. Because the correct boundaries between objects can be extracted by applying the proposed feature adjustment procedure and region merging in feature space, the problem of the false contours can be solved by the proposed approach.