LDSNet: A Lightweight Dynamic-Scale Network for Low-Light Salient Object Detection

Abstract

Low-light salient object detection (SOD) remains challenging due to degraded semantic cues, background interference, and scale variation under poor illumination. Existing methods either rely heavily on image enhancement or lack sufficient adaptation to illumination-dependent feature changes, which limits their robustness in complex low-light scenes. To address these issues, we propose LDSNet, a lightweight framework that integrates Global Context Compensation (GCC), Foreground-Background Decoupling Attention (FBDA), and Illumination-Guided Dynamic Scale Routing (IGDSR) in a unified pipeline. Specifically, GCC injects deep semantic priors into multi-scale features, while FBDA suppresses background-dominant responses and preserves target integrity. In addition, IGDSR jointly encodes a decoded-feature descriptor and a global luminance descriptor to generate input-conditioned routing weights for multi-scale receptive-field branches. Experiments on SOD-LLR show that LDSNet combines competitive accuracy with a compact model. Under the standard DUTS-TR protocol, a separately trained LDSNet model ranks among the top two listed methods on three normal-light benchmarks.