Comparative performance analysis of U-Net and DeepLabV3+ for semantic segmentation in traffic environments

Computer vision is an important field of artificial intelligence that enables machines to interpret and understand visual information from images. It is the basis of automated visual understanding in smart systems. Monitoring of streets, recognizing different vehicles and pedestrians, and comprehending the ever-changing traffic conditions for making right decisions are some of the features of modern intelligent traffic systems. It is indispensable for a computer to visually recognize the semantic segmentation (SS) of a scene, as it maps each and every pixel of the image to the categories like roads, vehicles, pedestrians, traffic lights, etc. However, existing traffic scene segmentation models often perform poorly under challenging conditions such as low lighting, blur, noise, and low resolution. These limitations limit the robustness of these methods. They can be a major obstacle to the development of autonomous vehicle technologies because they reduce the ability of perception systems to recognize a wide range of situations. This work proposes a new method for processing poor quality traffic images by sequentially applying super-resolution (SR), semantic segmentation (SS), and object detection with YOLOv8x. The SR module transforms degraded inputs, while U-Net and DeepLabV3+ are exploited for accurate pixel-level segmentation mask generation. Besides, YOLOv8x provides the precise object detection, which is really one of the critical steps for the avoidance of the errors in the crowded and complicated TSs. YOLOv8x uses bounding boxes to check segmentation and raise mAP. U-Net delivers PSNR of 41.93 dB, SSIM of 0.997, mIoU of 0. 750, and mAP of 0.950, whereas DeepLabV3+ yields PSNR of 46.03 dB, SSIM of 0.938, mIoU of 0.819, and mAP of 0.937.