Semantic Segmentation Network for Road Scenes Based on Improved U-Net

To address the deficiencies in traditional semantic segmentation models in extracting object edge information and accurately recognizing objects at various scales, this paper pro-poses a semantic segmentation network based on an enhanced U-Net architecture. Firstly, the original structure is substituted with a multi-scale modular network structure in the encoder section to bolster the network's feature extraction capability. Secondly, a comprehensive attention mechanism is integrated to achieve more precise feature extraction. Finally, a composite loss function, incorporating Dice Loss and Cross-Entropy Loss, is applied to thoroughly capture the features of localized regions in vehicle images. Experimental findings demonstrate a 10.20/0 improvement in mIo U accuracy on the Cityscapes dataset compared to the original U-Net model. The efficacy of the enhanced model in handling intricate scenarios, such as multi-scale targets and overlapping objects in road scenes, is duly validated.