Training a convolutional neural network on three-dimensional objects

This research examines an approach to training convolutional neural networks using three-dimensional computer graphics to represent a dataset. Converged neural networks are trained and classify objects using orthogonal projections of three-dimensional objects. One of the difficulties encountered in the classification process is object mirroring, which can be overcome by using Pearson correlation coefficient. This work analysed the architecture of existing solutions for efficient image recognition and visual information recording, which can be used as a basis for a convolutional neural network. To ensure the correct operation of the trained neural network, 3D graphics objects were created and after obtaining the required 3D model, black and white images were generated. It is on these black and white images that the object recognition software is trained. The successful functioning of the neural network has been experimentally proved. Thus, it becomes possible to recognize real objects based on convolutional neural networks trained in a virtual environment.