Tri-Branch GAN: A Semi-supervised Method Based on Rebalance

For deep learning applications in industrial scenarios, few-shot and imbalanced available datasets are very common. Traditional methods usually adopt the idea of hierarchical training, semi-supervised learning and the rebalancing method to train the model respectively, which has certain limitations: Separate trainings does not fully exploit the correlation between the two problems and causes additional computational overhead. Therefore, this paper proposes a semi-supervised learning method based on rebalance, named as Tri-branch GAN (Generative Adversarial Networks). This method makes full use of the correlation between the two problems, avoids the updating coating problem after the model parameter training, and saves the computational cost. Simulation results show that the proposed method can effectively improve the classification accuracy.