Using multi-resolution features and SOM in ultrasound images for segmentation

Texture boundaries or edges are useful information for segmenting a texture image. We present the experimental results of texture edge detection using 1-D and 2D multichannel, multiresolution gabor, log gabor and circular gabor filters and self organizing map (SOM). The traditional Gabor function is modified into a circular symmetric version to achieve rotation invariance. The log gabor provides the benefit of orthogonality and overcompleteness. SOM helps in reducing the dimension of a feature vector by producing a 1-D map which plots the similarities of data by grouping similar items together. The output of SOM is refined by post processing to obtain the texture edge map. The proposed methodology is tested on simulated as well as natural texture images and results are found to be satisfactory.