Abstract: To reproduce the ability of feature abstraction and self-learning of visual pathways, a feature perception model is proposed based on the hierarchical processing mechanism of optic nerve. On the basis of the improved convolution neural network, the model firstly established a retinal topological mapping in the primary visual feature processing according to the directional selectivity, the spatial locality and the lateral inhibition of neurons in retina; and then, provided a neural synaptic activation function to solve the problem of over-fitting in neural computation, by introducing the sparse representation of the biological nerve in the secondary visual feature processing; finally, the hierarchical visual perception computing model with computing perceptual invariance was proposed by reproducing information transmission of the ventral stream. The test results applying different scale data sets show that the proposed model has sound recognition effect on large scale target recognition problems, and the average recognition accuracy can reach 85%.