Dual stream Dual branch Network with Cascaded Codec Framework for RGB-D Saliency Detection

Depth maps have been proven to complement the problem of RGB images not having obvious features in salient object detection (SOD). Most existing RGB-D SOD models lack the ability to capture depth features and RGB feature spatial information during cross-modal fusion, which may lead to inaccurate fusion of features from significant regions. And we found that compared to low-level features, higher-level features have a greater contribution to performance, and improving the accuracy of higher-level features can significantly improve saliency detection performance. Therefore, in this article, we propose a cascaded coder-decoder dual stream dual branch network, which uses the initial saliency prediction obtained by attention branching to refine the high-level features of the backbone network. We also designed a Cross-modal space fusion (CSF) module to utilize the spatial information in RGB features and depth features to promote cross-modal fusion. A large number of experiments have shown that our model outperforms other RGB D SOD models on four widely used datasets.