Texture feature extraction via visual cortical channel modelling

A new algorithm is proposed for texture feature extraction and classification. The algorithm is based on the increasingly popular multichannel spatial filtering approach. A computationally convenient model is described for the hypothesized visual cortical channels. Each channel is tuned to a specific narrowband of spatial frequency and orientation, and is realized by two quadrature-phase Gabor filters which are intended to mimic an adjacent pair of simple cells. The means and the standard deviations of the channel output images are shown to be powerful texture features, and perform much better than the benchmark gray level co-occurrence matrix features under noise conditions