A Deep Learning Approach to Intelligent Detection of Shedthin Tile Pathology in Suzhou Classical Gardens

Abstract. Under the current rapid development of digitalization and artificial intelligence in heritage practice, deep-learning-driven pathology detection is emerging as a pivotal tool for preventive conservation. Deep learning-based intelligent pathology detection in heritage conservation has garnered increasing attention. This study explores intelligent pathology detection techniques using the YOLO11-seg model, taking the pathology identification of shedthin tiles of Suzhou classical gardens as a case study. Through data collection and annotation of 1,250 high-resolution images of shedthin tiles, a training dataset was constructed. After 362 training epochs, the model achieved automatic recognition of four key pathological type-water stains, color aberration, surface scaling, and excessive gaps-with respective accuracies of 79.31%, 73.38%, 61.12%, and 75.60%, and an overall accuracy of 74.38% that meets practical application requirements.that generally meets practical application requirements. The study further conducted quantitative analysis of detection results to assess the severity of shedthin tiles damage, providing critical references for formulating scientific restoration strategies. Compared with traditional visual surveys, the proposed workflow (i) increases detection speed by an order of magnitude, (ii) standardises assessments across inspectors, and (iii) captures early-stage micro-pathologies often overlooked in manual inspections. The results demonstrate that how integrating deep learning with heritage diagnostics can offer a replicable template for other fragile, repetition-rich surface historical materials of architectural heritage.