ETANet: An Efficient Triple-Attention Network for Salient Object Detection

Salient object detection (SOD) is a critical vision task in ubiquitous applications. Most existing methods have complicated structure and large number of parameters, which prevents these methods to deploy on practical applications. In order to solve this problem, we propose an efficient triple attention network (ETANet), which consists of multiple attention mechanisms. In detail, we design a crossed spatial-channel attention mechanism to extract useful low-level features, an efficient branch to perceive high-level features based on self-attention through multi-scale receptive field. In addition, we propose a dilated criss-cross fusion mechanism to fuse low-level and high-level features in an efficient way. The experiment results show that our architecture achieved competitive performance and can trade off between the accuracy and efficiency compared to other heavy-weight methods.