A Hierarchical Feature Map Model For Visual Object Recognition

The purpose of computer vision is to make computers recognize visualinformation like humans. Visual object recognition inspired by biologicalvisual information processing mechanisms is currently a research focus incomputer vision field. The main idea is to simulate the hierarchicalinformation processing from V1, V2, V4to IT visual cortex along the ventralpathway and then build mathematical models to recognize visual objects.Therefore, building computer vision models based on biological visualinformation processing mechanisms is of great significance and applicationprospect.In this paper, we first give a brief introduction on biological visualinformation processing mechanisms which includes the visual pathwaytheory and visual map theory, and then evaluate some of the hierarchicalmodels in computer vision based on these mechanisms. In order to learn thestructure of orientation map in the primary visual cortex (V1), we extend thelinear Reconstruction-ICA algorithm to its nonlinear form, combined withseveral features from other deep learning algorithms, and constrain the output of the algorithm to be smooth based on the presupposition on orientation mapby Hubel and Wiesel. The experiment result shows that our algorithm canlearn the2D structure of orientation map effectively.We also propose a new biologically inspired feed-forward hierarchicalmodel, i.e. the hierarchical feature map (HFM) model, for visual objectclassification. The four layered HFM model (Layer S, Layer C1, Layer C2and Layer C3) is based on the ventral pathway in visual cortex. Layer S useshand-craft Difference of Gaussians and Gabor patches to simulate theorientation map in primary visual cortex V1. Layer C1, C2and C3use radialbasis function (RBF) to simulate the receptive field (RF) property of eachnode and adopt competitive learning strategy to learn the exact shapes of theRFs based on visual map theory. We show that the proposed HFM model canwell preserve the main structure of the input image. We also demonstrate thatthe HFM model is capable of self-taught learning and abstract usefulrepresentation. The HFM model can achieve promising results on popularimage databases, which shows a great development prospect.