Deep learning-based image segmentation of urban scenes faces the problems of edge blurring and difficulty in distinguishing similar targets in practical applications. In this paper, an improved PSPNet model CBPSPNet incorporating CBAM attention mechanism is proposed to enhance the performance of urban scene segmentation by embedding a hybrid domain attention mechanism in PSPNet. The model combines the channel attention module and spatial attention module to adaptively focus on important features and suppress useless information. The experiments are validated with two datasets, Cityscapes and CamVid, using the SGD optimizer with the base learning rate set to 0.005 and power to 0.5, and 500 epochs of training. The results show that on the Cityscapes dataset, the CBPSPNet outperforms the traditional method on all evaluation metrics, with the range of evaluation metric values reaches 0.7-1, while the traditional method is only 0.6-0.9; it also exhibits faster convergence and lower loss values on the CamVid dataset. Ablation experiments demonstrate that using both average and maximum pooling together is more effective than using them individually. It is shown that the PSPNet model incorporating the CBAM attention mechanism can effectively improve the image segmentation accuracy of urban scenes.
