Unsupervised Texture Segmentation via Wavelet-based Locally Orderless Images (WLOIs) and SOM

M.K. Bashar, T. Matsumoto, Y. Takeuchi, H. Kudo, and N. Ohnishi (Japan)


texture, wavelet, locally orderless images , segmentation,self organizing map


Texture segmentation is a well-known problem in vi sual processing. Despite the existance of a number of techniques, reliable methods are only a few. Hence, we proposed a two step process through combining wavelet based local histogram features with Kohonen's self orga nizing maps(SOM). In the first step, we transformed the original image into orthogonal wavelet coefficients for non redundant representation of image-information. Then each discrete coefficient undergoes a non-linear transformation to obtain an isophote image, which is convolved with a spatial Gaussian to form a locally orderless image.We des ignate them as WLOIs, which specify a local histogram at each transform point. These GLOIs or statistical mo ments computed from them are the new texture features. In the second step, the WLOIs are applied to Kohonen's self organizing map for learning and segmentation. An experi ment with the standard Brodatz's texture database show the superior performance of WLOIs compared to conventional wavelet energies. Confusion matrix analyses confirm the above attributes.

Important Links:

Go Back