Training CNN Models for Welding Target Object Classification Using CAD-Based Synthetic Images

In certain industrial environments, it is difficult to collect many real images for specific target objects. To address this limitation and automate the classification process, an image classification model was developed using synthetic images generated through 3D modeling. In this study, Lego blocks were modeled in 3D CAD and rendered to produce photorealistic images. The rendered images were augmented using techniques such as flip and brightness adjustment to expand the dataset. Two convolutional neural network (CNN) models, ResNet18 and GhostNet, were trained and compared. A total of 8,960 augmented synthetic images and 1,000 real images were used for training and evaluation. As a result, ResNet18 showed a higher average classification accuracy than GhostNet. Consequently, ResNet18 was selected as the final classification model. In experiments comparing rendering styles, synthetic images with realistic shadows and materials achieved 95.6% accuracy, while less realistic images achieved only 86.0%. These findings suggest that the more visually realistic the synthetic images are, the more positively they affect the classification performance. Additionally, synthetic images with Gaussian noise achieved a classification accuracy of 99.2%, implying that incorporating noise characteristics commonly found in real environments can further enhance model performance.