A new image classification model based on brain parallel interaction mechanism

The brain vision mechanism provides a source for the structural design of convolution neural networks (CNNs). Inspired by the multi-stage interaction between the parallel ventral and dorsal stream of the human brain in the process of image recognition, we introduce a new image classification model called parallel interaction model (PIM). The feature extractor of PIM consists of two parallel CNNs, one of them is the main of the feature extractor connected to the classifier, and the other as a secondary of the feature extractor which can be used as a multi-stage interaction with the main one to help extract more effective features. Using the proposed PIM, we improve on two different CNNs, and validate model effects on Cifar-10, Aircrafts100 and Flower-17 data sets. We found that PIM produced significant performance improvements over the base networks. At last, before and after the interaction, the feature maps in one of the improved networks are visualized and analyzed.