Signal, Image and Video Processing Signal Processing

Multi-layer feature communication network for RGB-D salient object detection

Yan, Shiliang, Lu, Dandan, Sun, Tao, Wang, Yinling
Springer Nature · 2026-11-17 · DOI 10.1007/s11760-026-05794-5

Abstract

RGB-D salient object detection aims to identify and segment salient objects by exploiting complementary appearance and depth information. However, insufficient interaction between RGB and depth features can leave modality-specific noise unresolved, while inadequate communication across feature levels can hinder the integration of semantic cues and spatial details. To address these issues, we propose a Multi-layer Feature Communication Network (MFCNet), comprising a feedback optimization (FO) module and a multi-layer feature communication (MFC) decoder. The FO module integrates cross-modal information through channel and spatial attention and multi-scale context aggregation, and feeds the refined features back into both encoder streams. This feedback enables each stream to benefit from complementary cues in the other modality, helping suppress background interference and enhance salient regions. The MFC decoder aggregates high-level features to obtain an initial saliency prediction, then uses the resulting semantic guidance to refine low-level features while retaining their spatial details through residual connections. This design promotes consistent object localization and detailed saliency prediction through cross-level feature communication. Experiments on five public RGB-D benchmarks demonstrate competitive detection performance, while ablation studies validate the effectiveness of feedback optimization and guided cross-level refinement.

Read at publisher →