Frequency-based cross-attention fusion network for RGB-D salient object detection.

Abstract

In RGB-D salient object detection tasks, effectively fusing cross-modality information is crucial for improving detection performance. Current cross-modality fusion methods primarily focus on integrating the RGB and depth feature contents in the spatial domain, while the exploration of these fusion methods in the frequency domain remains relatively limited. To address this challenge, this paper proposes a Frequency-based Cross-Attention Fusion Network for RGB-D salient object detection (FCAFNet). Specifically, (1) a Frequency-based element-product Cross-Attention Module (FCAM) is employed to establish the long-range feature relationship between tokens from RGB and depth modality. (2) A Bi-directional Feature Aggregation Module (BFAM) is proposed to aggregate the horizontal and vertical feature contexts using different-size convolutional kernels. Finally, to refine the boundaries of salient objects, an edge supervision module (ESM) is proposed to sharpen the edge contour by the application of high- and low-frequency features in the bottom layers. Experimental results demonstrate that the proposed model outperforms the state-of-the-art models qualitatively and quantitatively. Our code is publicly available at: https://github.com/zhx-hub/FCAFNet.