Improve Neural network using Saliency

Abstract

In traditional neural networks for image classification, every input image pixel is treated the same way. However real human visual system tends pay more attention to what they really focus on. This paper proposed a novel saliency-based network architecture for image classification named Sal-Mask Connection. After learning raw feature maps from input images using a convolutional connection, we use the saliency data as a mask for the raw feature maps. By doing an element-by-element multiplication with the saliency data on the raw feature maps, corresponding enhanced feature maps are generated, which helps the network to filter information and to ignore noise. By this means we may simulate the real human vision system more appropriately and gain a better performance. In this paper, we prove this new architecture upon two common image classification benchmark networks, and we verify them on the STL-10 datasets. Experimental results show that this method outperforms the traditional CNNs.