Texture segmentation using fractal dimension and second order statistics

In this study, segmentation of textured images using four different textural features is examined. The first three features are fractal dimension (FD) of the original image, contrast-stretched image and top-hat transformed image, respectively. Contrast-stretching and top-hat transform are known as detail enhancement techniques in the presence of shading or poor illumination, thus it is assumed that the hidden structures in textures will be apparent after these transformations. The fourth feature, e.g. entropy, is one of the parameters estimated from spatial gray level co-occurence matrix statistics. For comparison purposes, two different feature smoothing methods are applied to the feature space before running k-ortalama clustering. The median smoothing gives more accurate segmentation results than EPNSQ (Edge Preserving Noise Smoothing Quadrant) approach. The experimental results are obtained by applying the proposed method on various natural texture mosaics. For mosaics of four textures the average segmetation accuracies are %96.8 and %96 for median smoothing and EPNSQ approach, respectively. The average segmentation accuracy for five textured mosaics is %95.5 with median smoothing, while it is %89 with EPNSQ approach. The experiments carried out with median smoothing for six and nine textured images give the segmentation accuracies as %94 and %92, while they are %84 and %87 with EPNSQ approach.