Symmetric Learning Data Augmentation Model for Underwater Target Noise Data Expansion

Abstract

An important problem in deep learning model training is the training of data. Without abundant data from real production environment for training, deep learning model would not be so popular as now. However, the cost to obtain abundant real data from real environment is great, especially from the underwater target environment. Therefore, it is feasible to generate more data similar to that in real environment. In this paper, a simple and easy symmetric learning data augmentation model (SLDAM) is proposed for underwater target noise data expansion and generation. The SLDAM, taking the optimal classifier of initial data set as discriminator, makes use of similar structure of classifier to construct symmetric generator based on antagonistic generation thought so as to generate data similar to initial data set to finish data expansion. This model has taken into consideration of loss function of feature loss and sample loss in the model training, which is able to reduce the dependence of generation and expansion on feature set. It has been verified by the experiment that SLDAM is able to achieve data expansion fast with a lower calculation complexity. By analyzing the experimental results, we know that SLDAM is able to generate abundant new data when it ensures the data recognition accuracy, then to construct a good production environment for application.