CDA-Net: Computer Vision based Automatic Car Damage Analysis

CDA-Net: Computer Vision based Automatic Car Damage Analysis

연구 분야: Artificial Intelligence

학회: ICVGIP '23: Proceedings of the Fourteenth Indian Conference on Computer Vision, Graphics and Image Processing

Car insurance claims are rising in tandem with the rising tide of car users. Every insurance claim requires an engineer’s manual assessment and a surveyor’s actual examination. This procedure can last anywhere from a few days to several weeks. Current deep-learning techniques have paved the way for this type of mechanization. Both the business and the client would benefit from a comprehensive system. To assess the damage cost for the insurance claim process, the make and model of the car, damaged parts, damage type, and severity of the damage are important parameters. We introduced two datasets, a piqued car make and model (CMM) dataset containing images of the most popular 23 car makes and 148 vehicle models available in the Indian automotive market. The second dataset consists of 11,380 images collected from insurance offices and web resources of different types of car damage, including annotations. In addition, it provides an assessment of the damage to each part, the severity of the damage (dents, scratches, bent, broken, cracks, smashed, punched, and pushed), and the location of the damage (front, back, side). The assessment helps estimate the damage cost when combined with structured data. The proposed CDA_YOLOv5 car damage assessment framework outperformed existing state-of-the-art one-stage models with an average per-class accuracy of 87.36% and an overall accuracy of 90.45%.