Analysis of the Efficiency of CNN Learning Based on Teacher-Student Training with Untrained Image Dataset
D.O. Zubarev, Inna Skarga-Bandurova · Bulletin of the National Technical University KhPI A series of Information and Modeling · 2018
Artificial neural networks expand the range of existing and potential spheres of use each year. The quality of training artificial neural networks is the basis of the quality of their further functioning. The article is devoted to the analysis of the effectiveness of the training of CNN class artificial neural networks for the recognition of an unprepared set of images (Image-Dataset) based on the principle of "teacher-student", in which the acting teacher of artificial neural network CNN‑1 pre-trained, which causes the learning algorithm, and the pupil An untrained artificial neural system CNN-2. It is proved that CNN-1 is more effective to search for a large spectrum of objects in images, and CNN-2 works best for narrowly focused precise searches of objects. Figs.: 7. Refs.: 22 titles.Keywords: artificial neural network; CNN; image; Image-Dataset.