The master's thesis examined the use of synthetically generated images to train deep learning models for object detection and segmentation in industrial contexts, where the shortage of high-quality annotated images is a significant obstacle. The developed software enables automatic generation of realistic images from CAD models, together with corresponding annotations, allowing customisation of scene diversity, complexity, and lighting conditions. The results indicate that models trained exclusively on synthetic data can learn basic object characteristics but have limited ability to generalise to real images, which improves significantly with additional training on a smaller set of real data. The study found that the diversity and realism of synthetic images greatly affect model transferability, while combining synthetic and real data is an effective approach for developing accurate and robust computer vision systems.
|
