Semantic segmentation of street view images based on improved U-Net model

This study proposes a street scene segmentation model based on attention and multi-level feature enhancement to address the problem of multi-scale object omission in traditional U-Net semantic segmentation of complex street scenes. Firstly, by integrating the scSE attention mechanism module, a collaborative mechanism of channel weight mapping and spatial saliency screening is established in the feature extraction stage to enhance the ability to focus on semantically sensitive areas; Then, an improved atrous pyramid module (ASPP) is embedded in the encoding and decoding architecture, utilizing parallel convolution with differential dilation rate to capture multi-level receptive field features and achieve dynamic fusion of local details and global semantics; Finally, by combining the mixed loss function of cross entropy and Dice coefficient, the gradient contributions of the majority and rare classes are balanced through adjustable weights to alleviate the prediction bias caused by long tail distribution. Experiments on the Cityscapes benchmark dataset showed that the proposed model MIOU improved by 3.93% compared to the original U-Net.