Problem

Conventional methods for inspecting car damage are labor-intensive, manual, and frequently fail to identify small surface flaws such as microscopic dents. These minor dents, despite their subtle appearance and low contrast, are critical for accurate insurance claims, quality control, and maintaining a vehicle's resale value. Existing automated systems and general vehicle damage datasets often prove inadequate for reliably detecting these subtle imperfections, highlighting a significant gap in current detection capabilities. A major challenge in this domain is the scarcity of high-quality, annotated datasets specifically focused on minute dents observed under varied real-world conditions.

Method

The authors developed a deep learning approach for automated minor dent detection, utilizing the YOLOv8 object recognition framework, specifically its YOLOv8m variant. To overcome the lack of suitable data, a novel, self-collected dataset named "Minor Dents Detection Dataset for Car Body Panels" was created. This dataset consists of 2,241 annotated images of car surfaces, capturing small dents under diverse lighting, angles, and textures, using an iPhone camera to ensure realism and relevance for real-world deployment. The dataset was structured into training (1,568 images), validation (449 images), and test (224 images) subsets, with bounding box annotations conforming to the YOLO format. Annotations were meticulously performed using the Roboflow platform, focusing exclusively on small, visible dents with tight bounding boxes, and a single class label was applied. Preprocessing steps included automatic image alignment via EXIF metadata and scaling all images to a 640x640 pixel resolution without preserving the original aspect ratio. Importantly, no artificial augmentations were added to the raw dataset; instead, real-time data augmentation techniques (e.g., random flips, scaling, rotations, HSV changes, mosaic, MixUp) were dynamically applied during the training phase to enhance robustness and generalization. Two models, YOLOv8m-t4 and YOLOv8m-t42, representing different training sessions with the same YOLOv8m architecture, were trained for 100 and 50 epochs, respectively, using an SGD optimizer and a cosine learning rate schedule. Model performance was evaluated using standard metrics including Precision, Recall, F1-Score, Intersection over Union (IoU), mean Average Precision (mAP@0.5), and Confusion Matrix analysis, alongside loss components like Box Loss, Class Loss, and Distribution Focal Loss.

Results

The evaluation demonstrated that the YOLOv8m-t42 model generally surpassed YOLOv8m-t4 in performance. Specifically, YOLOv8m-t42 achieved a precision of 0.86, a recall of 0.84, and an F1-score of 0.85, with its mAP@0.5 stabilizing at 0.60 and a PR curve area of 0.88. This model recorded 431 true positives and 366 false positives. In comparison, YOLOv8m-t4 yielded a precision of 0.81, a recall of 0.79, and an F1-score of 0.80, with a mAP@0.5 of 0.624 and a PR curve area of 0.82. It identified 432 true positives but had a higher number of false positives at 424. While YOLOv8m-t4 showed slightly faster convergence and one more true positive detection, YOLOv8m-t42 exhibited superior precision (fewer false alarms) and better overall generalization capability. Analysis of training loss trends revealed that both models effectively reduced losses, although YOLOv8m-t4 showed signs of overfitting after approximately 30 epochs, as indicated by its validation loss curve. Interpretability was enhanced through visualizations of activation maps and predicted bounding boxes.

Implications

The proposed deep learning method offers a scalable and cost-effective solution for accurate identification of minor vehicle dents, addressing a critical need in the automotive industry. Its high detection accuracy and low inference latency make it well-suited for real-time applications such as automated insurance assessments, quality control, and intelligent maintenance systems. The creation of a custom, real-world dataset provides a valuable resource for further research, bridging a gap in publicly available datasets for subtle dent detection. Future work could focus on developing lightweight models optimized for edge device deployment, integrating 3D input data for more comprehensive surface analysis, and incorporating advanced attention mechanisms to further improve generalization and reliability, particularly in challenging environments with reflections and varied lighting conditions. Current limitations identified include sensitivity to reflections and lighting, as well as the need for more diverse data for unusual dent types.