Evaluation of the effectiveness of a data set expansion method based on deep reinforcement learning
Abstract
Evaluation of the effectiveness of a data set expansion method based on deep reinforcement learning
Incoming article date: 17.12.2024The article presents the results of a numerical experiment comparing the accuracy of neural network recognition of objects in images using various types of data set extensions. It describes the need to expand data sets using adaptive approaches in order to minimize the use of image transformations that may reduce the accuracy of object recognition. The author considers such approaches to data set expansion as random and automatic augmentation, as they are common, as well as the developed method of adaptive data set expansion using a reinforcement learning algorithm. The algorithms of operation of each of the approaches, their advantages and disadvantages of the methods are given. The work and main parameters of the developed method of expanding the dataset using the Deep-Q-Network algorithm are described from the point of view of the algorithm and the main module of the software package. Attention is being paid to one of the machine learning approaches, namely reinforcement learning. The application of a neural network for approximating the Q-function and updating it in the learning process, which is based on the developed method, is described. The experimental results show the advantage of using data set expansion using a reinforcement learning algorithm using the example of the Squeezenet v1.1 classification model. The comparison of recognition accuracy using data set expansion methods was carried out using the same parameters of a neural network classifier with and without the use of pre-trained weights. Thus, the increase in accuracy in comparison with other methods varies from 2.91% to 6.635%.
Keywords: dataset, extension, neural network models, classification, image transformation, data replacement