SR-SegDrive: Enhancing UNet-Based Semantic Segmentation for Autonomous Driving with Super-Resolution GAN

For autonomous driving systems, accurate perception and segmentation rely on high-quality visual inputs. However, low resolution inputs are often a result of poor lighting or technological constraints in real-world scenarios, which hinders segmentation performance. Super Resolution Generative Adversarial Network (SR-GAN) and a U-Net based segmentation model are combined in this study to create SR-SegDrive, a two stage system. In damaged photos, SRGAN recovers high-frequency information to enhance UNet's semantic segmentation of lane lines and drivable zones. The benefits of resolution augmentation in driving scenarios are highlighted by the proposed method's notable improvements in mean Intersection over Union (mIoU) and pixel accuracy when tested on the BDD100K dataset.