Formation of artificial data from inert materials for training computer vision neural networks using the example of crushed stone production

The article presents a research approach to the formation of synthetic data for training neural networks of computer vision systems using the technological process of crushed stone production as an example. The main attention is paid to the problem of scarcity and heterogeneity of real images used in the training of industrial computer vision models. In order to overcome these limitations, a method for generating photorealistic images based on the Unity game engine and the Unity Perception toolkit is proposed. An algorithm for generating pseudorandom objects based on the deformation of the icosphere using Perlin noise has been developed, which made it possible to form variable three-dimensional models of crushed stone of various shapes and fractions. The created system provides productivity of up to 1000 images per minute and automatic creation of annotations in YOLO format. The results of training the YOLO v11 model on the generated synthetic sample demonstrated high quality indicators: mAP50 = 0.96 and mAP50-95 = 0.90, which confirms the effectiveness of the proposed approach. The developed system can be used to generate training samples in conditions of limited access to real data, as well as adapted to various applied tasks of industrial computer vision.