Open Access

Lukas Lins

• Programmable Logic Controller (PLC) modules are used in industrial settings to control and monitor various manufacturing processes. Detecting these modules can be helpful during installation and maintenance. However, the limited availability of real annotated images to train an object detector poses a challenge. This thesis aims to research object detection of these modules on real images by using synthetic data during training. The synthetic images are generated from CAD models and improved with Generative Adversarial Networks (GANs). The CAD models are rendered in different scenes, and perfectly annotated images are automatically saved. A technique called domain randomization is applied during rendering. It renders the modules in different poses with constantly changing backgrounds. As the CAD models do not visually resemble the real modules, it is necessary to improve the synthetic images. This project researches StarGAN and CycleGAN for the task of image-to-image translation. A GAN is trained with real and synthetic images and can then translate between these domains. YOLOv8 and Faster R-CNN are tested for object detection. The best mean Average Precision (mAP) is achieved when training with a synthetic dataset where 50% of the images were improved with StarGAN. When trained with YOLOv8 and evaluated on a real dataset, it achieves a mAP of 84.4%. Overall, the accuracy depends on the quality of the CAD models. Using a GAN improves the detection rate for all modules, but especially for unrealistic CAD models.