Object Recognition Based on Biologic Visual Mechanisms

For speedy robust object recognition, our model builds on Serre’s standard model and modifies it by additional biological characteristics, such as introducing the manner of neuron firing, feature localization, and merging unit features in the higher layers. According to the four-layer architecture of standard model, we first apply Gabor filters on a higher intermediate frequency band of original image, and then compute the number of firing neurons in the next layer to create shift- and scale-tolerance. To build up feature complexity we use prototype matching for all scales in the third layer. In the last layer we max pool local units to achieve larger shift-tolerance. We use SVM at the classifying stage. Tested on the Caltech datasets, our improved model offers a significant gain in speed with fewer features and achieves even better state-of-the-art performance than standard model.