Car damaged detection and evaluation

Abstract

The proposed car damage detection system follows a multi-stage cascade approach, where each stage consists of a classifier and a bounding box regressor. The system leverages the Cascade Mask R-CNN architecture with a Swin-FPN backbone, which provides multi-scale contextual information for accurate and robust object detection. The system is trained on a large dataset of labeled car images, including various types of car damages such as dents, scratches, cracks..., to learn the discriminative features for car damage detection. To evaluate the performance of the proposed system, extensive experiments are conducted on a benchmark dataset of car images with ground-truth car damage annotations. The results show that the Cascade Mask R-CNN-based car damage detection system achieves good performance in terms of detection accuracy and computational efficiency. The proposed system is able to accurately localize and segment car damages in images