Active Deep Learning Technique for Underwater Target Recognition

Underwater target recognition task based on sonar signal data suffers from both professionalism requirement and high time cost during data annotation stages. After deep learning methods are introduced into the field, the quantity and quality of training data have become one of the key factors affecting the model performance. In order to solve the contradiction between data requirements and data annotation cost, this paper introduces active learning techniques, selecting more valuable and performance-improving data from unlabeled data pool using specific data selection strategies. An active deep learning framework that iteratively integrates active learning methods to the training process of deep learning models is proposed. Experiments on real sonar signal data show that compared to random selecting method, the prediction accuracy of the model increases faster with the growth of annotated data volume when using active learning algorithms. It takes only 51
^~
52% of the total training data to achieve the same model performance when all the data is used for training. Effectiveness of the proposed technique on unbalanced dataset and under noisy conditions is also tested, which verifies its robustness in complex marine environment.