Encoding Co-occurrence of Features in the HMAX Model


We introduce a method for encoding co-occurrence of features in the HMAX model of visual recognition, and conduct a series of experiments to investigate the contribution of co-occurrence towards better recognition performance. We show that classification accuracy is increased by adding a higher-order layer to the HMAX processing hierarchy, whereby co-occurrence of features is encoded as a new dictionary of features. We show that concatenation of mean pooling, max pooling and co-occurrence information results in better classification results on three datasets (Caltech101, a subset of Caltech256, and TMSI Underwater Images). Overall, we show that incorporating co-occurrence statistics into a biologically-inspired model of visual recognition provides a boost in classification performance above that produced by incorporating occurrence statistics alone.

Back to Table of Contents