Computer vision with AI can improve the production and quality control process by enabling the early detection of defects in products. This has significant implications for the automotive industry, as it can help reduce costs, minimize waste, and ensure that products are of the highest quality. However, deploying AI-based computer vision systems is often challenging because it requires a large amount of data that may not be available, especially for new products. This is because it's necessary to have data representing all possible defects and consider all variability in environmental factors such as lighting, position, orientation, and noise. To overcome the challenge of requiring a large amount of data for AI-based computer vision control in manufacturing, we propose a novel approach in this paper that utilizes synthetic data for training. Our methodology is based on creating a virtual environment using Unity software which involves a five-stage process: First, (1) we upload a 3D model of the product and the quality inspection environment. Next, (2) we apply features that represent different possible defect classes such as missing, damaging, scratch, etc. Then, (3) we add Physics-Based Rendering (PBR) materials to the 3D model to reduce the gap between real and synthetic data. After that, (4) we apply randomizing features such as position, rotation, scale, lighting, and post-processing to achieve a randomized environment. Finally, (5) we render the environment using the High-Definition Render Pipeline (HDRP) to generate thousands of labeled images of the product. Our approach offers several advantages over traditional methods, including the ability to generate common defects in manufacturing, the ability to configure the environment to improve detection robustness, and the ability to automatically label each image, to be aligned with the detection algorithms used. However, we acknowledge that the method presents some challenges, such as difficulty in preparing the 3D model, the optimization of the mesh, issues when applying materials to the 3D model, or the impossibility of applying some defect features to small parts. Despite these challenges, we used the generated dataset to train AI models for detecting manufacturing and assembly defects using one-stage detectors like YOLO and two-stage detectors like R-CNN. The results were positive and demonstrated the effectiveness of using synthetic data instead of real images, which are sometimes unavailable and consume a lot of time and resources. In the paper, we present two successful use cases of our method. The first use case involved the detection of defects in a simple part produced by a stamping machine with limited defect features. The second use case involved a more complex part produced on an assembly line with wires and many plastic or metallic components, with a significant number of possible defects. In both cases, our approach showed excellent results, we were able to detect defects, which prove the potential of synthetic data generation for training AI models. Moreover, we demonstrate that a combination of synthetic images and real images during the training stage can improve the AI algorithm's accuracy. This can help overcome some of the limitations of synthetic data, such as the inability to capture all possible variations in the real world. The combination of synthetic and real data can also help improve the detection of rare defects that may not be present in the synthetic dataset. We conclude that synthetic data generation is a valuable tool for implementing AI modules in production lines in the automotive industry. The ability to generate synthetic data can help overcome the challenge of data unavailability and improve the effectiveness of the AI-based computer vision system. Our future work aims to automate the pipeline of dataset generation so that mechanical or process engineers can complete it in a few hours, which is essential in the automotive industry where product references or production conditions can change rapidly. By doing so, we hope to further improve the efficiency and effectiveness of AI-based computer vision systems in the manufacturing industry.
Mr. khalid Kouiss, Senior Expert Automation, Forvia