Soft computing approach for color image segmentation and texture classification
Ajoy Kumar Ray, Somnath Sengupta, Milind M. Mushrif · 2010
Color image segmentation and texture classification are challenging tasks in image analysis. This thesis presents novel techniques for color image segmentation and texture classification using soft computing methodologies. The mathematical foundation of Histon has been presented here. The histon is a contour plotted on the top of the histograms of the primary color components of a color image. It exploits the correlation among the neighboring pixels in the same plane as well as the other color planes. The concept of roughness index has been introduced to correlate the histogram and the histon. The roughness index plotted against the intensity, exhibits crests and troughs similar to the histogram with well defined peak and valley points. The proposed color thresholding algorithm based on the histon roughness index, yields considerable improvements in the segmentation performance when compared with other conventional techniques. The concept of histon has been extended to fuzzy histon, where the belongingness of a pixel to the similar color sphere is not binary, and is decided using a Gaussian membership function. To improve the performance of the algorithm further, we have introduced the concept of A-IFS histon using the Atanassov’s intuitionistic fuzzy set (A-IFS) theory. A novel technique for A-IFS representation of image has been developed to deal with the inherent hesitancy in deciding the nature of the pixel, edge or non-edge. A quantitative study to measure the quality of segmentation using probabilistic Rand index has been carried out. The results demonstrated on the Berkeley segmentation Database reveal the effectiveness of the proposed algorithm. A new distance measure based on A-IFS has been proposed to deal with the missing information in the data in the form of hesitancy in an intuitive way. The proposed distance measures are employed to develop a new A-IFS homogeneity histogram model for color image segmentation. The effectiveness of the proposed distance measures has been compared with the existing of intuitionistic fuzzy distance measures and the results have been demonstrated in a color image segmentation application. The problem of texture classification has also been investigated in this thesis. There is an inherent limitation of the conventional wavelets in the analysis of the high frequency signals with relatively narrow bandwidth. Also, the non-orthogonality and high computational cost of the Gabor wavelets sometimes limits their applicability. An alternate feature extraction scheme using cosine-modulated wavelets, which occupy adjacent bands in the spectrum, have better frequency resolution and low design and implementation cost, has been suggested. Finally, we proposed a novel soft-set theory based classifier with high classification rates and low computational complexity. The proposed technique compares favorably well with other feature extraction and classification techniques.