Abstract

Salient object detection (SOD) in optical remote sensing images (ORSIs) remains challenging under low-light conditions, where reduced illumination impairs object visibility, contrast, and boundary clarity. Existing methods, primarily designed for well-lit scenes, suffer significant performance degradation in such scenarios. To address this, we propose DarkSalNet, a dual-branch framework tailored for low-light ORSISOD. It jointly exploits raw low-light inputs and UVI-enhanced features—derived from a perceptually guided color space transformation—to capture complementary global and structural cues. Two cross-modal integration modules are introduced: (1) a CrossModal Fusion (CMF) module that applies bidirectional gating for adaptive spatial calibration, and (2) a Semantic CrossAggregation (SCA) module that employs cross-attention and dual-path refinement to enhance semantic consistency. Extensive experiments on six synthetic low-light ORSI-SOD datasets and one real low-light NSI-SOD dataset demonstrate that DarkSalNet consistently outperforms state-of-the-art methods across varying illumination levels.

Datasets

The synthetic dataset will be made available after the publication.

Resources

🔗 Pretrained Weights

• Download DarkSalNet Weights (Baidu)

🔗 Saliency Maps

• Download Saliency Maps (Baidu)

---

Model train

Download the pre-trained model weights and dataset. Modify the dataset path in the config file or training script. Run training:

python train.py

Model evaluation

Download the trained LWMNet weights. Set the weight paths in inference.py. Run inference:

python inference.py

Results