SCD-IGNet: Enhanced Semantic Segmentation of Low-Light Rainy Images via Symmetric Cross-Decoupling and Illumination Guidance

Semantic segmentation, a critical research area in computer vision, is essential for various applications, including autonomous driving and intelligent transportation. However, existing semantic segmentation methods often perform poorly in low-light rainy conditions due to challenges such as underexposure, occlusion, and rain streak noise. To address these issues, we propose a novel semantic segmentation method for low-light rainy images, named SCD-IGNet. This method leverages symmetric cross-decoupling and illumination guidance to enhance segmentation accuracy. SCD-IGNet decouples low-light rainy images into content and style components using a decoupling enhancement module, which comprises a global and local feature extraction module and a differential content enhancement module. The content components are then processed by a backbone network to extract features, while the style components generate illumination features that aid in semantic segmentation. Experimental results demonstrate that our method achieves a 1.1% improvement in mean intersection over union and a 2.1% increase in mean accuracy compared to the baseline method on the NightCity-rain dataset. Additionally, our method exhibits excellent performance on the NightCity-fine dataset, validating its efficacy for low-light rainy image semantic segmentation. The code is available at https://github.com/Liusir765832/SCD-IGNet.git.