Data Augmentation using Conditional Generative Adversarial Network for Underwater Target Recognition

In underwater target recognition tasks, a deep learning recognition model is more effective than traditional methods, but meanwhile requires sufficient labeled data at the same time. Data augmentation is an effective method to increase the size of training data and reduced the mismatch between training and test. Different from the traditional approaches by directly adding noise to the original waveform or perform waveform transformation, we explore a supervised conditional generative adversarial network framework for data augmentation by using few ship data. The experiment result shows that higher recognition accuracy can be obtained when generated data are added to the original data. We find that the recognition models do not always perform better when the added generated data increases. It is appropriate to add about 5 times of the amount of the original data. The experimental result shows that in 5-target recognition task, with only 1000 samples for each type of target provided, a relative 8.8% to 13.3% improvement can be obtained when adding the generated data.