This study introduces a novel application of state-of-the-art object detection models for automating quality control in automotive manufacturing, presenting the first comprehensive comparative analysis of YOLOv8, YOLOv9, and YOLOv10 architectures for vehicle damage detection. Utilizing a custom-curated dataset of 7,258 images, we employ transfer learning techniques to optimize model performance, a pioneering approach in this domain. Our results demonstrate the unprecedented superiority of YOLOv10 across key metrics, achieving a mean Average Precision (mAP50) of 0.65077 and an F1-score of 0.64934. We uniquely quantify the effectiveness of transfer learning, showing substantial performance gains with pre-trained weights initialization. Notably, we establish YOLOv10's viability for real-time quality control applications despite marginally increased computational requirements, a finding not previously reported. This research contributes novel insights into AI-driven solutions for automotive quality control, advancing the digital transformation of manufacturing processes and paving the way for future industrial AI innovations.
The integration of computer vision systems in manufacturing processes has become crucial for
automated quality control, especially in the automotive industry. Traditional manual inspection
methods are time-consuming, subjective, and error-prone, necessitating more robust and efficient
quality control mechanisms [
Recent advancements in deep learning, particularly the YOLO (You Only Look Once) family of
models, have shown promise in real-time object detection [
Transfer learning has emerged as a powerful technique, allowing pre-trained models to be
finetuned for specific tasks with limited domain-specific data [
Previous studies have demonstrated the potential of deep learning in manufacturing quality
control [
The implementation of an effective automated quality control system has far-reaching
implications for the automotive industry, potentially increasing throughput, improving
consistency, and reducing waste [
Our study explores how advanced object detection models and transfer learning can enhance automated quality control for vehicle damage detection. This research contributes to computer vision and manufacturing technology, offering insights for industry practitioners adopting AIdriven quality control solutions.
Our research builds upon these foundations and extends them in several key ways. Firstly, we
provide a comprehensive comparison of the latest YOLO models, building on the work of Redmon
and Farhadi who introduced YOLOv3 [
One of the key challenges in implementing computer vision systems for quality control in
manufacturing is the need for large, diverse, and accurately labeled datasets. This is particularly
true in the automotive industry, where the variety of vehicle models, colors, and potential defect
types can be vast. Transfer learning offers a promising solution to this challenge by allowing
models pre-trained on large, general datasets to be fine-tuned for specific tasks with relatively
small amounts of domain-specific data [
The effectiveness of transfer learning demonstrated in this study aligns with the comprehensive survey by Tan et al., who provided an in-depth overview of deep transfer learning techniques and their applications [16]. Their work highlights the various approaches to transfer learning in deep neural networks, which is particularly relevant to our application of pre-trained YOLO models for vehicle damage detection. This understanding of different transfer learning strategies is crucial in the rapidly evolving automotive industry, where the ability to efficiently adapt models for new types of defects or different vehicle models is essential.
This research addresses real-world implementation challenges [17], aligning with Industry 4.0 principles [18-19], and extends beyond defect detection to broader manufacturing process optimization [20-22].
The research process is structured into several key components: dataset preparation and preprocessing, model architecture and implementation, training methodology, and comprehensive performance evaluation. Through this systematic approach, we aim to provide insights into the effectiveness of these advanced computer vision techniques for enhancing quality control processes in the automotive industry [23-26].
We utilized the Car Dents Computer Vision Project dataset, comprising 7,258 images (6,855 training, 377 validation, 26 test) of various vehicle damage types. Preprocessing steps included: 1. Resizing images to 640x640 pixels. 2. Data augmentation: 90° rotation, ±15° shear transformation, and ±15% brightness adjustment.
Dataset analysis revealed class imbalance (Dent: 3391, Accident: 1927, Scratch: 2072) and bounding box characteristics, informing model optimization strategies.
We implemented YOLOv8, YOLOv9, and YOLOv10, leveraging their architectural advancements: by: 1. YOLOv8. Anchor-free detection, new backbone network. 2. YOLOv9. Efficient neck structure, advanced loss functions. 3. YOLOv10. Dynamic attention mechanism, hybrid backbone (convolutional and transformer layers).
Transfer learning was applied using pre-trained weights from the COCO dataset, represented θnew=θ pre+ ∆θ , where θnew are the new model parameters after fine-tuning;θ pre are the pre-trained parameters; ∆θ represents the parameter updates during fine-tuning.
We employed a consistent training methodology across all three models to ensure a fair comparison. The key aspects of our training process were: 1. Optimizer. We used the Adam optimizer with a cosine learning rate schedule. The learning rate can be described by the equation: π lr (t )=lrmin+0.5 ∙( lrmax−lrmin )∙( 1+cos ( t ∙ )) ,
T
Ltotal= λ1 ∙ Lbox + λ2 ∙ Lobj+ λ3 ∙ Lcls , where t is the current epoch; T is the total number of epochs; lrmin is the minimum learning rate; lrmax is the maximum learning rate.
2. Loss Function. We utilized a combination of losses typical for object detection tasks: where Lbox is the bounding box regression loss; Lobj is the objectness loss; Lcls is the classification loss; λ1, λ2, and λ3 are weighting factors. (1) (2) (3) 3. Early Stopping. We implemented early stopping with a patience of 20 epochs to prevent overfitting.
Performance assessment utilized: Mean Average Precision (mAP@0.5 and mAP@0.5:0.95), Precision and Recall, F1-Score, Confusion Matrix, and Inference Time
The precision and recall can be calculated using the following equations: where TP is True Positives; FP is False Positives; FN is False Negatives.
The F1-score is then calculated as:
Precision=
Recall=
TP TP + FP
TP , TP + FN
, F 1= 2 ∙ Precision ∙ Recall ,
Precision+ Recall (4) (5) where Precision represents the proportion of true positive predictions among all positive predictions made by the model; Recall (also known as sensitivity or true positive rate) indicates the proportion of true positive predictions among all actual positives in the dataset.
Experiments were conducted using PyTorch on NVIDIA GeForce RTX 4080 Super GPUs, implementing models via the Ultralytics YOLO framework. The procedure included model initialization, training, validation, testing, and performance analysis.
To ensure reproducibility, a fixed random seed was used across all experiments, allowing fair comparisons between YOLO versions while controlling for neural network training stochasticity.
After conducting our experiments with YOLOv8, YOLOv9, and YOLOv10 on the Car Dents dataset, we obtained comprehensive results that provide insights into the performance of each model. In this section, we will present and analyze these results in detail.
All models showed consistent improvement during training, with YOLOv10 exhibiting the fastest convergence. YOLOv10 achieved the lowest final box loss (1.1842) and classification loss (0.80735), significantly outperforming YOLOv8 and YOLOv9. YOLOv10 demonstrated the highest mAP50(B) throughout training, peaking at 0.65077.
Models successfully detected various damage types with high confidence (0.3-0.9) and demonstrated robustness to diverse scenarios. Multiple damages on single vehicles were effectively identified. Some misclassifications were observed, particularly between dents and scratches.
YOLOv8
YOLOv10 consistently outperformed other models across all metrics, with a 5.7% improvement in F1-score over YOLOv8 and 3.1% over YOLOv9.
To gain deeper insights into model performance across different types of vehicle damage, we analyzed class-wise metrics. Figure 1 presents the F1-Confidence curves for each class (Accident, Dent, Scratch) across all three models.
YOLOv10 achieved the highest F1-scores for all damage types: dents (0.719), scratches (0.689), and accidents (0.637). Accident detection proved most challenging across all models.
Figure 2 illustrates the Precision-Recall curves for each model, providing a comprehensive view of their performance across different confidence thresholds.
YOLOv10 demonstrated the largest Area Under the Curve (AUC) and maintained higher recall in high precision regions (>0.8) compared to YOLOv8 and YOLOv9.
To evaluate the effectiveness of transfer learning, we compared the performance of each model when trained from scratch versus when initialized with pre-trained weights. Table 2 presents this comparison:
These results demonstrate the significant benefits of transfer learning across all models. Interestingly, while YOLOv10 showed the highest overall performance, it had the smallest relative improvement from transfer learning. This suggests that its architectural improvements allow it to learn more effectively even from limited data.
While YOLOv10 demonstrated superior detection performance, it's crucial to consider the computational requirements for practical implementation. Table 3 compares the model sizes and average inference times:
The marginal increase in model size and inference time for YOLOv10 is relatively small compared to the performance gains, suggesting that it remains a viable option for real-time applications in manufacturing settings.
Our comprehensive study on implementing computer vision systems for automated quality control in automotive manufacturing, focusing on vehicle damage detection, has yielded significant insights into the capabilities of state-of-the-art object detection models.
YOLOv10 consistently outperformed its predecessors, achieving a mAP50 of 0.65077 and an F1score of 0.64934, representing improvements of 5.7% and 3.1% over YOLOv8 and YOLOv9, respectively. The application of transfer learning proved highly beneficial, with YOLOv8, YOLOv9, and YOLOv10 showing mAP50 improvements of 22.5%, 20.4%, and 16.9% respectively when initialized with pre-trained weights.
Despite its superior performance, YOLOv10 required only marginally more computational resources, with a negligible increase in inference time (14.1ms compared to 12.5ms for YOLOv8), making it viable for real-time applications in manufacturing settings.
Key implications for the automotive manufacturing industry include: 1. Enhanced Quality Control. Automation of damage detection can reduce human error and increase consistency. 2. Increased Efficiency. Real-time defect detection enables inspection without production bottlenecks. 3. Cost Reduction. Minimizing manual inspection and early defect detection can lead to significant cost savings. 4. Adaptability. Transfer learning enables quick adaptation to new defect types or vehicle models. 5. Data-Driven Insights. Deployment of these systems can generate valuable data on defect patterns and trends.
This study demonstrates the significant potential of advanced object detection models, particularly YOLOv10, in revolutionizing quality control processes in automotive manufacturing. The success of transfer learning techniques paves the way for widespread adoption of AI-driven solutions in industrial quality control, contributing to enhanced product quality and manufacturing efficiency.
The authors have not employed any Generative AI tools.