A Curated Dataset and Deep Learning Approach for Minor Dent Detection in Vehicles

Muhammad Danish Zia Baig, Mohsin Kamal · 2025

Read A Curated Dataset and Deep Learning Approach for Minor Dent Detection in Vehicles by Muhammad Danish Zia Baig, Mohsin Kamal (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-15431.txt.

Introduction

Muhammad Danish Zia Baig and co-authors describe a deep learning system, published in 2025, that finds tiny dents on car bodies. Manual checks are slow and often miss small flaws, so the team built a fresh image set for training. Their photos show car surfaces under varied light, angles, and textures, with each dent marked.

The team trained the YOLOv8m object detector and two tuned variants on this set. The best variant, YOLOv8m-t42, reached a precision of 0.86 in spotting the small flaws. It also ran fast enough for live use. The authors point to uses in automated insurance checks and routine car inspection, where small damage is easy to miss by eye.

Catalog record

CREATORMuhammad Danish Zia Baig, Mohsin Kamal
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-15431

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Curated Dataset and Deep Learning Approach for Minor Dent Detection in Vehicles Muhammad Danish Zia Baig, Mohsin Kamal Abstract Conventional car damage inspection techniques are labor-intensive, manual, and frequently overlook tiny surface imperfections like microscopic dents. Machine learning provides an innovative solution to the increasing demand for quicker and more precise inspection methods. The paper uses the YOLOv8 object recognition framework to provide a deep learning-based solution for automatically detecting microscopic surface flaws, notably tiny dents, on car exteriors. Traditional automotive damage inspection procedures are manual, time-consuming, and frequently unreliable at detecting tiny flaws. To solve this, a bespoke dataset containing annotated photos of car surfaces under various lighting circumstances, angles, and textures was created. To improve robustness, the YOLOv8m model and its customized variants, YOLOv8m-t4 and YOLOv8m-t42, were trained employing real-time data augmentation approaches. Experimental results show that the technique has excellent detection accuracy and low inference latency, making it suited for real-time applications such as automated insurance evaluations and automobile inspections. Evaluation parameters such as mean Average Precision (mAP), precision, recall, and F1-score verified the model’s efficacy. With a precision of 0.86, recall of 0.84, and F1-score of 0.85, the YOLOv8m-t42 model outperformed the YOLOv8m-t4 model (precision: 0.81, recall: 0.79, F1-score: 0.80) in identifying microscopic surface defects. With a little reduced [email protected]:0.95 of 0.20, the [email protected] for YOLOv8m-t42 stabilized at 0.60. Furthermore, YOLOv8m-t42’s PR curve area was 0.88, suggesting more consistent performance than YOLOv8m-t4 (0.82). YOLOv8m-t42 has greater accuracy and is more appropriate for practical dent detection applications, even though its convergence is slower. The visualizations of activation maps and anticipated bounding boxes further improve interpretability. The suggested method provides a scalable and economical solution for minor dent identification, laying the groundwork for future research into 3D surface analysis and mobile platform deployment. Keywords Minor Dent Detection \cdot YOLOv8 \cdot Deep Learning \cdot Object Detection \cdot Real-Time Inspection \cdot Automotive Damage \cdot Computer Vision 1 Introduction Minor dents are one of the most commonly missed but serious types of surface damage found during a vehicle examination. These dents, which are frequently shallow and irregular in appearance, pose a unique challenge for both human evaluators and automated systems because to their low contrast with adjacent surfaces and subtle geometric deformations. Detecting faults in vehicles is crucial for insurance claims, quality control, and after-sales inspections, as even minor flaws can impact resale value and customer satisfaction Thomas et al. 2023 Raju et al. 2023 . Traditionally, this activity has been conducted manually, relying on experienced specialists whose assessments might be subjective and inconsistent, particularly under changing lighting and surface conditions. To address this demand, deep learning-based object identification models, particularly those from the YOLO (You Only Look Once) family, have proven to be exceptionally effective. YOLO models are particularly valued for their ability to strike a balance between detection accuracy and speed, making them ideal for real-time applications in industrial and commercial environments. The most recent version, YOLOv8, offers significant improvements in detection speed, model correctness, and flexibility in a wide range of complex circumstances. A recent study S. et al. 2025 employed YOLOv8 to build a real-time vehicle damage detection system, as demonstrated in Figure. 1 . The dataset used to train this system classified eight different types of automobile damage, including shattered windshields, dents, and damaged bumpers. This foundation provides an excellent starting point for more specific research into identifying smaller types of damage, such as small dents, which requires even more accurate feature extraction and classification. Figure 1: System Architecture of Car Damage Detection S. et al. 2025 . Recent improvements in deep learning have resulted in highly specialized models that can detect even minor and sparse surface flaws. Zhang et al. 2023 improved the YOLOv5 architecture by incorporating a BiFPN-based feature fusion technique and an extra small-object detection layer, resulting in significant gains in detecting minute faults on metallic surfaces. Although their study is not vehicle-specific, the technical issues they addressed are similar to those found in automobile dent detection. Similarly, Agrawal et al. 2025 developed a YOLOv8-based hybrid model that can categorize and segment numerous damage types shown in Figure. 2 , including dents, across several vehicle sections, highlighting the versatility of such architectures. Figure 2: System Architecture of Car Damage Detection S. et al. 2025 . Transfer learning has also gained popularity in this field. Sruthy et al. 2021 fine-tuned CNN models such as ResNet50 and VGG16 using labeled vehicle damage datasets. Their approach enabled pre-trained models to adapt to specific visual patterns of car damage, such as tiny dents. They did, however, notice constraints caused by dataset imbalance and insufficient segmentation granularity, both of which are common concerns with dent detection methods. These findings highlight the need for more refined models and diverse training datasets that can generalize across a variety of environmental conditions. To meet real-time processing constraints, researchers have optimized lighter variants of YOLO, such as YOLOv7-tiny and YOLOv7x for automotive damage detection Ramazhan et al. 2023 . When trained on annotated photos of scratches, dents, and shattered glass, these variations achieved strong F1-scores, with YOLOv7x reaching 0.949. The use of image augmentation techniques, which improved robustness to lighting and orientation variability, increased the effectiveness of such models even more. However, their ability to detect microscopic, low-contrast dents remains strongly dependent on dataset quality and augmentation approach. A growing amount of research has begun to use segmentation and detection approaches for greater performance in difficult circumstances. Bhatlawande et al. 2023 introduced a hybrid pipeline that uses YOLOv5 for object localization and watershed segmentation for contour extraction. This method allows for fine-grained recognition of dent shapes, making it ideal for shallow and mild surface deformations. Other studies, such as those by Wang et al. 2023a and Chua et al. 2021 , have introduced methods such as structured light imaging and modular CNN-based classifiers that identify distortions on reflective surfaces and distinguish specific auto parts and their damage status. Several recent studies have attempted to address the difficulty of dent detection using a variety of methodologies. Polyantseva et al. 2025 developed a method for localizing deformations using Laplacian-based preprocessing and unsupervised K-Means clustering, which proved effective in controlled situations. However, their method failed with generalization under different illumination and background textures. Widjojo et al. 2022 used deep CNN-based approaches, specifically localized patch learning and transfer learning, to greatly enhance the classification of visible dents. While these models performed admirably in detecting obvious surface defects, their sensitivity to tiny, shallow, or shaded dents remained a shortcoming. Collectively, these findings highlight the need for more universal and resilient architectures—such as YOLOv8 and transformer-based models—that can capture minute visual cues across a wide range of real-world scenarios. Despite these developments, a substantial barrier persists: a lack of high-quality, annotated datasets focusing on tiny dents. Generic datasets for vehicle damage identification, such as CarDD Wang et al. 2023b and VehiDE Huynh et al. 2023 , frequently have few examples of subtle faults. This absence of representation reduces model generalization and training effectiveness. Researchers such as Widjojo et al. 2022 and Polyantseva et al. 2025 advocate for domain-specific dataset preparation, including multi-view and augmented samples, to increase performance on less visible damage types. In light of these

Read the complete text