Saliency Detection Framework Based on Deep Enhanced Attention Network

Abstract

In recent years, RGB-D based salient object detection has received increasing attention. Most of the previous models are based on early fusion and late fusion to fuse the input RGB-D images. Such an approach cannot effectively explore the complementary information on RGB-D images, and thus information loss and incomplete fusion can occur. To address these problems, we propose a novel framework for RGB-D salient object detection (DEANet). Our framework uses Siamese networks to extract features of RGB-D images to mine the similarity between the two modal data onto a shared-weight manner. Specifically, we propose the Channel-Spatial Fusion Block (CSF) for feature condensation of the extracted Depth features, which can capture the spatial and channel dependencies on RGB-D images from both spatial and dimensional perspectives, thus better mining the complementary information between the two modal images and enables the thoroughness and completeness of the features for which useful information is extracted. In addition, we propose an edge-optimization loss (EL) to obtain smoother salient object edges by supervising the edges of the objects. Comprehensive experiments on four popular evaluation metrics show that our DEANet is able to generate RGB-D saliency detectors with high robustness and generalization ability. In particular, DEANet outperforms thirteen current state-of-the-art methods in terms of four evaluation metrics on six challenging datasets.