Research of regularization techniques for SAR target recognition using deep CNN models

Abstract

Due to limited training data for SAR target recognition tasks, different regularization techniques have been used in the deep convolution neural networks (CNN) to improve generalization ability. In this paper, the influences of three regularization techniques, including data augmentation, L2 regularization term, dropout, are studied under standard operating conditions (SOC) when moving and stationary target recognition (MSTAR) dataset is used for SAR target recognition. Four representative CNN models based on classical models, such as AlexNet and ResNet, are selected and trained to recognize 10-classes targets. Additionally, a CNN model which has fewer network parameters is designed based on multi-scale spatial feature extraction strategy and SqueezeNet to study the influence of the amount of network parameters. The experimental results demonstrate that, when using the AlexNet series model for SAR target recognition, using dropout may greatly improve the ability of model optimization. ResNet series models which have more layers, have better effect on Test 1+noise than other CNN models, especially taking dropout in the model. For the models based on highway networks, adding L2 regularization terms in loss function can improve the test accuracy, but it also makes the latter phase of training extremely unstable. Data augmentation is an effective regularization technique when the model can get high training accuracy.