Efficient Road Segmentation Techniques with Attention-Enhanced Conditional GANs

Road segmentation from aerial images is a challenging yet crucial task, underpinning significant applications in urban planning, navigation, and transportation systems. In this study, we employ a conditional Generative Adversarial Network (GAN) architecture that synergistically integrates the strengths of Attention U-Net and PatchGAN to address this task. The Attention U-Net, serving as the generator, is trained on the publicly available Massachusetts Roads dataset, with an emphasis on critical regions while concurrently disregarding the irrelevant ones, thereby enhancing the accuracy of road segmentation. Simultaneously, the PatchGAN discriminator ensures the generation of sharp, high-quality segmentations. Through this cooperative approach, we have achieved an overall accuracy of 98.2%, a recall of 82.30%, a precision of 78.66%, an Intersection over Union (IoU) of 67.19%, and an F1 score of 80.44% on our dataset. While these results are promising, they also highlight areas for improvement, particularly in reducing false positives and enhancing the identification of all road pixels, underscoring the potential for future refinements in this research domain.