In the dynamic field of autonomous driving, the study explores the crucial yet underexplored realm of night scene semantic segmentation. Despite advancements in daytime scenarios, challenges persist in nighttime scenes due to insufficient exposure and a scarcity of labeled data. To address these issues, the NightCity dataset was enriched by creating labeled nighttime images with diverse low-light conditions. Techniques such as adjusting illumination, contrast, brightness, and introducing blur were employed to enhance the dataset’s complexity. This augmented dataset was then used to train an existing model, incorporating Dual Image-Adaptive Learnable Filters (DIAL-Filters). Comprising a Context-Aware Image Processing Module (CAIPM) and a Learnable Guided Filter (LGF), the DIAL-Filters aim to enhance semantic segmentation under varying illuminations during nighttime driving conditions. The CAIPM module integrates differentiable image filters within a compact neural network, adaptively improving images for enhanced segmentation under diverse illuminations. The Learnable Guided Filter (LGF) also refines the segmentation network output, contributing to the final segmentation result. The model underwent end-to-end hybrid supervised training using both the augmented and Cityscapes datasets, producing noteworthy results in night scene semantic segmentation. This dual-module approach optimizes the model for effective semantic segmentation in various nighttime driving conditions.