Abstract:

A new feed forward hierarchy image classification method is proposed, which is based on the latest neurophysiological findings about cortex recognition and local feature method. The proposed method concentrates on feature specificity and invariance. First, the normalized input image is decomposed by convolving the image with multi-scale and multi-orientation Cabor filters to extract the corresponding elementary tuning features. Then sparse features are obtained by Center-Symmetric Local Binary Pattern (CS-LBP) algorithm. Internal templates and global max pooling operation are used to get scale invariance and position invariance. The hierarchy feature processing outperforms pure local feature representation, like Scale Invariant Feature Transform (SIFT), in image classification experiment. Moreover, the proposed method is competitive with current computer vision algorithm on Caltech 256 object categories. The results strengthen the case for using this bionic method in image classification and more general visual category appliance.

Key words: computer application, image classification, object recognition, local feature, visual cortex