Corners as interesting points in biologically inspired object recognition, HMAX

In this paper a new approach is proposed to improve accuracy, robustness and process time in HMAX for object recognition. The HMAX is a hierarchal biologically inspired model which leads to a good performance in object recognition. Despite achieving a relatively good classification rate, its result is not stable, and it is varied during each program run, which means it is not a repeatable approach. Using randomly selected features, the HMAX has an inconstant classification rate. We propose to change the strategy of feature selection in the standard HMAX. By repeatable feature selecting, the HMAX achieves a very good repeatable performance which is more reliable in comparison with the previous result. To cope with unrepeatability in the HMAX, we suggest that corners which are extracted by the Harris corner detection can be selected as key points. By this alternation, we receive a higher classification rate and a lower computation time. The proposed approach shows excellent performance especially when the number of training images and extracted features is low. In the training stage, only five images for positive classes and five images for negative classes are used. Classification rate and time consumption are evaluated in Caltech dataset. Furthermore, the effect of the number of feature is demonstrated in both new approach and the standard HMAX features.