Natural Image Recognition Baded On Human Visual Model

The visual system of human is the best object recognition system at present, which can extract informations from the complicated surroundings easily. Many researchers tryed to join the mechanisms of the visual system of biology with computer vision and built a model that accorded with the process of visual information, which provided a new way for the development of computer vision. Reseachers proposed many visual models presently, most of which cann't simulate the information processing of visual system. The standard model proposed by Serre is succeed in recognition, and we had some changes on this model in order to get a better visual system.We introduced the base characteristic about visual cortex and a few of models which were successful firstly. We improved the standard model based-on the dual-tree complex wavelet which is approximate shift invariant and had the lower redundancy. In the new model, there were some progress in the recognition rate and computational speed. Combined with the sparse features proposed by Jim Mutch, the redundance would be reduced.Second, another new visual model-EAM was also proposed based on the standard model. This model was improved by replacing the max operation with the mechanism of neural activation. Combined with Marr-like pyramid which is scale-invariant and shift-invariant, the complex wavelet, local features, the model could get the features that were local invariant.The third, we proposed a feedback model which used the log-gabor wavelet that accorded with the visual characteristic better and the top-down attentional mechanism. The top-down attentional mechanism could get rid of the noisy informations in the surroundings and centralized the features that belong to the objects. At last, we compared the improved models with the standard model in recognition tasks on Caltech dataset. The results showed that the improved models had some progress in the recognition rate and computational speed.