CAE-Net: Cross-Modal Attention Enhancement Network for RGB-T Salient Object Detection

RGB salient object detection (SOD) performs poorly in low-contrast and complex background scenes. Fortunately, the thermal infrared image can capture heat distribution of scenes as complementary information to image, so RGB-T SOD has recently attracted more attention. Many researchers have committed accelerating development SOD, but some problems still remain be solved. For example, defective sample interfering contained or hinder model from learning proper saliency features, meanwhile low-level features with noisy result incomplete objects false positive detection. To solve these problems, we design a cross-modal attention enhancement network (CAE-Net). First, concretely fusion (CMF) module fuse where cross-attention unit (CAU) is employed enhance two modal channel used dynamically weigh features. Then, joint-modality decoder (JMD) cross-level are purified by higher level multi-scale sufficiently integrated. Besides, add single-modality (SMD) branches preserve modality-specific information. Finally, employ multi-stream (MSF) three decoders’ Comprehensive experiments conducted on datasets, results show that our CAE-Net comparable other methods.