2019 IEEE International Scientific-Practical Conference Problems of Infocommunications, Science and Technology (PIC S&T) | 2019
Deep Learning Neural Networks with Using Pseudo-Random 3D Patterns Generator
Abstract
The paper presents the method of generating the learning patterns for the artificial neural networks which purpose is to recognizing and detecting objects in the digital images. The main objective of presented method is to create learning patterns by random deformations and modifications of recognized 3-dimensional model of objects. The paper describes algorithm for creating learning series of patterns and reference database to it, architecture of used artificial neural network model, briefly environment Unreal Engine 4 used to program samples/patterns generator, the learning algorithm and modification of backpropagation algorithm improving quality and time of neural network learning. The main part of this algorithm is presented in the form of block diagrams and appropriate math equations. The next chapter describes the effectiveness of neural network learning based on the patterns generated only on the 3-dimensional models. The effectiveness was estimated in dependence of used neural architecture and activation function. The experiment results are presented in tables and appropriate to them figures. At the end of the paper are final conclusions concerning obtained results of calculations.