A Deep Learning-Based Method for Visual Element Recognition and Openness Evaluation of Street Interfaces

To enhance the intelligent perception and quantifiable analysis of urban street spaces, this paper proposes a street interface openness evaluation method that integrates deep learning-based semantic segmentation with visual metric modeling. An improved DeepLabv3+ semantic segmentation network is constructed, using ResNet-101 as the backbone, with the incorporation of atrous convolutions and an ASPP module to achieve multi-scale feature extraction, thereby improving the recognition accuracy of typical street scene elements such as buildings, sky, and greenery. Based on the semantic segmentation results, an openness calculation formula is designed, which performs weighted fusion based on the pixel proportions of different visual elements to achieve a quantitative description of street permeability. The experimental section constructs a self-collected street scene image dataset and designs ablation and comparative experiments to comprehensively validate the contribution of each module to performance. The results demonstrate that the proposed model outperforms traditional models in key metrics such as mIoU, PA, and openness error, with the lowest openness error reaching 6.317%. This study provides an efficient and interpretable technical solution for urban design and street space evaluation, showing promising practical application prospects.