Contrastive Learning Image Enhancement for SOD in Haze

Deep learning-based saliency object detection (SOD) methods have achieved promising results on standard datasets. However, these algorithms often struggle with low-quality images captured under hazy conditions. While image recovery-based approaches can enhance image quality, they typically rely on synthetic data, limiting their effectiveness in real-world scenarios. To address this limitation, this paper proposes an end-toend framework that integrates an adaptive image enhancement module with SOD. Specifically, we design a Multi-Differentiable Image Processing (MDIP) module capable of adaptive image enhancement. Additionally, we introduce contrastive learning to ensure that the enhanced image aligns more closely with clear images in the feature space, thereby improving performance on real-world datasets. Experimental results demonstrate that the proposed framework significantly outperforms state-of-the-art SOD methods on both synthetic and real-world hazy datasets.