A Segmentation Method of Color Texture Image

Texture analysis has long been an important research area of image comprehension and computer vision. While the ability of human to distinguish different textures is apparent, the automated description, recognition, and segmentation of these same patterns have proven to be quite complex. Most existing texture segmentation approaches focus on gray texture images, while this paper presents a color texture image segmentation method, which is based on fractal theory and BP neural network. There are two important aspects of texture image segmentation procedures, namely feature extraction and segmentation. Most texture features often being used in segmentation are based on such models as structural model, statistic model, filter model, random field model and fractal model. The texture features studied in this paper are all based on the fractal theory. This choice is motivated by the observation that the fractal dimension (FD) is relatively insensitive to an image scaling, and shows a strong correlation with human judgement of surface roughness. But fractal dimension alone does not provide sufficient information to describe and segment natural textures. Fractal sets may share the same fractal dimension and yet have strikingly different appearances or textures. Thus multifractal theory is introduced in the paper. The texture feature based on multifractal theory is spectrum function D(q), which can characterize efficiently natural textures even if they are quite similar. A new class of texture measures based on the concept of lacunarity is defined. The second-order statistic lacunarity is small when the texture is dense, and large when the texture is coarse. Based on the above work, this paper presents a color texture image segmentation method. This approach converts color texture image from RGB format to HSI format and computes fractal dimension, multifractal function q-D(q), and lacunarity by intensity as texture features. Normalized hue and saturation are used as other classification features. BP neural network is adopted as classifier.The experiment of segmenting the color texture images has been done and the result is satisfactory, which verifies the effect of this method.

Перевод пока недоступен