Visual Saliency Prediction Using Attention-based Cross-modal Integration Network in RGB-D Images

Abstract

Saliency prediction has recently gained a large number of attention for the sake of the rapid development of deep neural networks in computer vision tasks. However, there are still dilemmas that need to be addressed. In this paper, we design a visual saliency prediction model using attention-based cross-model integration strategies in RGB-D images. Unlike other symmetric feature extraction networks, we exploit asymmetric networks to effectively extract depth features as the complementary information of RGB information. Then we propose attention modules to integrate cross-modal feature information and emphasize the feature representation of salient regions, meanwhile neglect the surrounding unimportant pixels, so as to reduce the lost of channel details during the feature extraction. Moreover, we contribute successive dilated convolution modules to reduce training parameters and to attain multi-scale reception fields by using dilated convolution layers, also, the successive dilated convolution modules can promote the interaction of two complementary information. Finally, we build the decoder process to explore the continuity and attributes of different levels of enhanced features by gradually concatenating outputs of proposed modules and obtaining final high-quality saliency prediction maps. Experimental results on two widely-agreed datasets demonstrate that our model outperforms than other six state-of-the-art saliency models according to four measure metrics.