Learning complex cell features with cooperating pooling operation for object recognition

Abstract

A simple biologically inspired feature extraction algorithm is proposed for object recognition. First, a set of statistical topographic filters modelling the properties of complex cells in a primary visual cortex (V1) are learned based on enhanced independent subspace analysis (EISA), and locally invariant feature maps are extracted by convolving the filters with each image. Then, the cooperating cortical pooling operations which combine the energy model and the MAX-like model are used to increase the phase and shift invariance of the filter response. Experimental results on the MNIST dataset and the Caltech101 dataset demonstrate that the algorithm is efficient and achieves high recognition accuracy.