RC部材の地震損傷画像からのひび割れセグメントの自動抽出とデータの構造化手法の提案

抄録

This study presents an integrated method that combines deep-learning models and image-processing algorithms to detect cracks in RC member damage images while distinguishing crossings, discontinuities, and branches. Cracks are parameterized by a luminance-based polyline approximation; widths at approximation nodes are classified by a trained model, and lengths are measured from nodes. Damage images from full-scale RC partial-wall lateral-loading tests served for training and evaluation. The image-based lengths achieved an average ratio of 0.99 relative to manual measurements, and width classification accuracy was 0.88. Findings indicate crack segments with quantitative length and width attributes can be extracted.