Unsupervised texture classification of images using cortex filters

摘要

We present a block-based multi-channel mechanism for unsupervised texture classification of images inspired by Human Visual System (HVS). The proposed approach compresses the large feature space by logical selection of block We employ 2D Gaussian functions, regarded as cortex filters, to simulate the band pass nature of simple cells in HVS. Within the frequency plane of each data block, filters are defined in various radial bands and orientations and used to obtain a set of feature images whereby texture features are defined by computing average energy. The obtained feature images are thus integrated with 'k-means clustering' for unsupervised classification of homogeneous textural regions. We demonstrate our method by several experiments on real world and synthetic images. Confusion matrix analysis projects the superiority of our method con pared to gray level co-occurrence matrix (GLCM) approach.