Adaptive Data Augmentation Training Method for SAR Military Target Classification

Synthetic Aperture Radar, as a primary means of military reconnaissance, necessitates its related automatic identification technology to possess sufficient reliability and portability. However, the application of neural networks for the identification of target radar images often encounters a severe issue of overfitting. In this study, we propose an adaptive data augmentation technique that gradually expands the training set. Initially, we apply various data augmentation methods to the images in the dataset, in which there is a crucial denoising method proposed by us. Subsequently, the augmented images are trained alongside the original images, enabling the network to autonomously learn the correlations between the images, thus achieving the objective of reducing overfitting. Our experiments, conducted by comparing the training of multiple models on the FAST and MSTAR datasets, demonstrate the excellent performance of the adaptive data augmentation method in enhancing model attention and combating overfitting. We achieve up to maximum 10% accuracy improvements on FAST datasets with only changes in data enhancement