Investigation of Methods for Increasing the Efficiency of Convolutional Neural Networks in Identifying Tennis Players

Abstract

Abstract The article is devoted to the study of the effectiveness of the convolutional neural networks (CNNs) application for solving the problem of tennis players face recognition. For ease of analysis, two players were selected: Roger Federer (Switzerland) and Rafael Nadal (Spain). To isolate faces from the publicly available images of the players, it is proposed to use the Haar cascades and the Viola–Jones method. These images are used to train and test convolutional networks with various parameters: architecture, including the number of layers; epochs of learning; optimization methods; and also when applying various regularization methods, including drop out and data augmentation. The use of regularization made it possible to reduce the effect of overfitting. In addition, the efficiency of networks with pretrained layers based on transfer learning methods is investigated. The VGG-16 convolutional network is chosen for the transfer learning. For a large number of different combinations of convolutional networks, metrics are calculated for precision, recall, and accuracy. The average gain for these parameters is 25% with the best set of characteristics for convolutional networks and training. It is also shown in the study that the patterns of applying certain modifications are universal for optical images. In particular, similar architectures and training approaches are also tested for the problem of recognizing cats and dogs on a much larger dataset. The study confirms the average increase in recognition metrics of 26%.