Old Residential Building Defect Inspection Based on Hierarchical Labeling Metric Learning

Old residential areas, which have been inhabited for many years and lack professional property management, frequently encounter issues such as deteriorating infrastructure and unauthorized alterations by residents, resulting in numerous safety hazards. This study collects the Old Residential Area Inspection Dataset and identifies building safety inspection as an integration of multiple complex fine-grained image classification tasks. Current models usually only focus on specific types of fine classification. To overcome this limitation, a hierarchical label metric learning approach is proposed. This method leverages metric learning to bring features of samples from the same parent class closer together while pushing those from different parent classes farther apart. The cross-entropy loss function is simultaneously applied to ensure accurate classification at the child-class level. The proposed approach improves top-1 and top-5 accuracy on ORAID by over 1.4% compared to models trained with only cross-entropy loss. Significant performance gains are also observed on the CIFAR-100 and tieredImageNet datasets, which similarly possess hierarchical label structures.