Deep-Layer Training of CNN for SAR with Two-Stage Data Augmentation

This study explores a novel two-stage data augmentation technique to expand the training capacity of the small synthetic SAMPLE dataset through the mechanism of deep layer training. Ultimately, the present investigation seeks to improve the classification performance of Convolutional Neural Networks (CNN) when trained on synthetic SAMPLE data and tested against measured SAR imagery. One possible explanation for the non-optimal performance of CNN in such scenarios is detectable high-level differences between synthetic and measured SAR imagery. When fine detail information is removed from synthetic datasets, the augmented data can train deep layers more readily. By also changing background texture, overfitting is reduced. Combined, these augmentations improve a CNN's effectiveness at classifying measured data. To achieve this two-stage augmentation, a combination of kernel operations, threshold-based segmentation, and alpha blending are used. Comparative evaluations using transfer learning and multi-resolution datasets are performed with the five-layer AConvNet CNN. Both test accuracy and training activity are evaluated to provide qualitative evidence of deep layer training. Improved accuracy in these experiments provides quantitative evidence that this two-stage data augmentation can effectively improve the training capacity of synthetic SAR imagery.