Combined saliency enhancement based on fully convolutional network

Abstract

We propose a combined saliency enhancement architecture by combining two traditional saliency enhancement strategies: saliency aggregation and saliency optimization. Previous methods have presented many remarkable saliency maps. Saliency aggregation fuses these results to highlight the salient objects and suppress the background. Saliency optimization optimizes the rough computational saliency maps by local and global context in the original image. We first illustrate the principle of saliency aggregation and optimization, and how to implement these two strategies using fully convolutional network. And then, we propose a network based on FCN to combine these two strategies. We use FCN to iteratively combine the results of the two strategies. Our method is evaluated on five representative datasets. Experimental results indicate that our architecture outperforms the state-of-the-art methods.