Diffusion-based domain-anchored refinement for vehicle dent detection enhancement

초록

Vehicle dent detection is important for artificial intelligence-based vehicle inspection in insurance claim assessment, rental and car-sharing return inspection, and used-car condition reporting. However, defect-free vehicle images are relatively easy to collect, whereas dent-positive images are scarce and costly to annotate. Unlike scratches or contamination, dents are impact-induced geometric deformations whose appearance is mainly governed by illumination- and viewpoint-dependent shading rather than stable local textures. These properties make reliable dent detection difficult because training data are often limited and imbalanced, while dent appearances vary substantially with illumination and viewpoint. In addition, depending on their location, dents may also indicate functional or assembly-related risks, including panel misalignment or module displacement. To address these challenges, this study proposes Diffusion-Based Domain-Anchored Refinement, a generative data augmentation method for training vehicle dent detectors. The proposed method generates dent patterns using box-based patches, refines them through domain-anchored image-conditioned reverse diffusion, and overlays the refined patches onto clean vehicle surfaces with automatic bounding-box labeling. Experiments on a customer-built dataset and the Car Damage Dataset show that the proposed approach reduces the Fréchet Inception Distance by 22% and the Kernel Inception Distance by 38% on average compared with existing surface-defect generation methods. For downstream dent detection, the generated data improve mean average precision at an intersection-over-union threshold of 0.50 by 2.47% and mean average precision averaged over thresholds from 0.50 to 0.95 by 3.12%. These results support scalable and cost-effective vehicle dent inspection through practical training-time augmentation. © 2026 Elsevier Ltd.

키워드