Saliency Prediction with Relation-Aware Global Attention Module

Abstract

The deep learning method has achieved great success in saliency prediction task. Like depth and depth, the attention mechanism has been proved to be effective in enhancing the performance of Convolutional Neural Network (CNNs) in many studies. In this paper, we propose a new architecture that combines encoder-decoder architecture, multi-level integration, relation-aware global attention module. The encoder-decoder architecture is the main structure to extract deeper features. The multi-level integration constructs an asymmetric path that avoid information loss. The Relation-aware Global Attention module is used to enhance the network both channel-wise and spatial-wise. The architecture is trained and tested on SALICON 2017 benchmark and obtain competitive results compared with related research.