Abstract
<title>Abstract</title> <p>RGB-D salient object detection leverages RGB images and depth maps to identify visually conspicuous regions. Despite progress, two challenges persist: depth maps often contain noise or missing data, and high-precision models suffer from excessive parameters and complexity. This paper presents RGBA-Net, a lightweight network with an asymmetric dual-stream encoder using EdgeNeXt and MobileNetV3 for RGB and depth features. A Depth Reliability Gate mitigates low-quality depth inputs via gradient-guided calibration. Cross-modality interaction is achieved through a decoupled dual-branch structure, while a Multi-scale Boundary Enhancement Fusion Module refines boundaries via adaptive fusion and progressive decoding. Experiments on seven datasets show that RGBA-Net outperforms twelve mainstream methods with only 3.49 million parameters. It sets new state-of-the-art benchmarks for lightweight models, confirming its balance between efficiency and accuracy.</p>