A swin transformer-based asymmetrical network for light field salient object detection
Yunqing Li, Jiquan Ma · 2023
In this paper, we propose a Swin Transformer-based Asymmetrical Network (SwinAN) for light field salient object detection (SOD). SwinAN is a bifurcated backbone network composed of a 2D branch and a 3D branch. The 2D branch is driven by Swin Transformer blocks for all-in-focus images, while the 3D branch is constructed by a 3D convolutional neural network (CNN) for focus stacks. Then, the high-level features extracted by the asymmetric backbone network are integrated using cascaded Multi-Level feature Fusion (MLF) modules, which consist of channel attention mechanisms and a residual block. Finally, the Multi-Attention Fusion (MAF) module is used to fuse multi-modal features to generate the predicted saliency maps. Experimental results show that after comparing 10 different SOD models, our SwinAN achieves excellent performance on 3 datasets, confirming the superiority as an efficient and accurate method for detecting saliency