Extracting urban spatial perception attributes and scene elements by integrating VGG-16 and CBAM

Abstract As urbanization continues to accelerate, there is a growing need for the analysis of urban spatial perception attributes and scene elements. In response, the research proposed a multi-scale perception network-based model for extracting urban scene elements and an attention-enhanced segmentation network-based model for analyzing urban spatial scene structures. The urban scene feature extraction model incorporated Siamese convolutional neural networks and convolutional block attention to achieve multi-scale perception extraction. The urban spatial scene structure analysis model combined a dynamic attention module with an encoder-decoder architecture to enhance the accuracy of scene element segmentation. During testing of the urban scene feature extraction model at a resolution of 768, its classification accuracy and cross entropy were 95.4% and 0.065, respectively. The model's average ranking accuracy for beauty, comfort, and cleanliness was 92.5%, 91.8%, and 93.2%. In testing the urban spatial scene structure analysis model, the boundary intersection to union ratio and boundary F1 score were 81.2% and 82.1%, respectively, with a boundary complexity of 0.6. The results demonstrated that the proposed method excelled in tasks such as perceptual attribute classification and scene element parsing, effectively addressing complex and diverse urban spatial features.