A multiresolution texture gradient method for unsupervised segmentation

摘要

A texture segmentation algorithm is described, which is designed using a cooperative algorithm within the Multiresolution Fourier Transform (MFT) (Wilson et al., IEEE Trans. Inform. Theory 38(2) (1992) 674–690) framework. The magnitude spectrum of the MFT is employed as feature space wherein the detection of texture boundaries is estimated by means of the combination of boundary information and region properties. A pre-smoothing process is first applied to the MFT magnitudes to reduce the texture fluctuations followed by Sobel operator performed on the MFT magnitudes to give a texture boundary estimate. The refinement of the boundary estimate is accomplished by utilizing both boundary link probabilities and region link probability in an iterative manner. Results on a number of synthetic and natural textures that illustrate the effectiveness of the scheme are presented.