Сreating image datasets for training neural networks of сonveyor video analytics

The paper presents a technology for generating synthetic data for training computer vision neural networks used in industrial quality control tasks. A method of image generation using the Unity game engine is proposed, which makes it possible to create photorealistic scenes of the production process, simulate the movement of objects on the conveyor and form various types of laminate defects. Algorithms for changing textures, lighting, camera angles, and defect characteristics are implemented, which ensures high diversity and realism of the training sample. To test the effectiveness of the approach, the YOLOv11 model was trained on the generated data. The results obtained (mAP50 = 0.95) confirmed the ability of the model trained on synthetic images to adequately process real data and identify defects on the production line.