A CNN Based Method to Solve Class Imbalance Problem in SAR Image Ship Target Recognition

Abstract

Synthetic Aperture Radar (SAR) recognition algorithms have made great progress with the use of convolutional neural networks (CNNs). However, the imbalanced distribution of ship targets causes CNNs trained by conventional methods to fail to recognize rare ship types in the marine target recognition field. Traditional resampling methods have disadvantages of overfitting and low recognition accuracy, perform poorly in SAR images. To address this problem, this paper proposes a dynamic weighted sampling and soft threshold moving combined method which can be used in most current CNN based SAR recognition algorithms. Specifically, dynamic weighted sampling changes training data distribution according to the validation set performance, and soft threshold moving adjusts output decision thresholds of trained CNNs. The test results on the public dataset OpenSARShip show that our method can effectively identify rare ship types while obtaining a higher recognition accuracy than traditional resampling methods.