EPDPM-SinGAN: Enhancing urban street semantic segmentation with region-wise GANs feature

Abstract

Real-time semantic segmentation is essential in various applications, including autonomous driving and urban scene comprehension. This research focuses on the integration of discriminative and generative models to provide real-time semantic segmentation in complicated urban landscapes. This study introduces a novel model called EPDPM-SinGAN, which utilizes a SinGAN to extract context-aware features, together with an AdvVGG16 encoder and a U-Net decoder. The technique amplifies edge and texture characteristics to address typical challenges in semantic segmentation, particularly occlusions, and variations in object sizes. We incorporate Hierarchical Attention Mechanisms with Adaptive Feature Fusion to enhance the segmentation process and prioritize informative features. Finally, the Secondary Discriminative Pixel Mining (SDPM) module is introduced to target informative pixels for refined segmentation in complex urban scenarios. Our proposed technique EPDPM-SinGAN outperforms other segmentation models on the Cityscapes and CamVid datasets by achieving mIoU of 81.27 % and 78.7 % respectively, establishing itself as the current state-of-the-art.