Application of visual recognition technology for housing characteristic in urban physical examination

Abstract With the rapid advancement of urbanization in China, urban challenges such as environmental pollution, ecological degradation, and public health risks have become increasingly prominent. As a critical diagnostic tool for urban development issues, the Urban Physical Examination (UPE) mechanism has been widely adopted across multiple cities. However, the current UPE system faces significant challenges in data checking: (1) the identification of problematic images relies on manual review, where accuracy is vulnerable to staff’s expertise and subjective judgment, leading to inconsistent evaluations; (2) traditional manual screening is inefficient and hard to meet the growing demand for urban health assessments. To enhance both the accuracy and efficiency of issue identification, this study introduces automated image recognition technology to assist the manual inspection process. Focusing on three common UPE indicators—missing surveillance cameras at unit entrances, absent signage at unit entrances, and damaged stair treads. This study utilizes the EfficientNet-B0 convolutional neural network model, comparing three training approaches: fine-tuning of fully connected layers, fine-tuning all layers, and training the model from scratch. Experimental results demonstrate that EfficientNet-B0 achieves outstanding performance across multiple indicators in the housing dimension of UPE, with all accuracy rates exceeding 88.6%. Notably, the accuracy for the stair tread damage/intact indicator surpasses 98%, indicating exceptional discriminative capability. Notably, the total training time for the model was merely 1.15 hours and testing in 0.6 hours, which is only 3% of the manual review time. These outcomes provide a powerful technical means for improving urban management efficiency and safety evaluations.