Speckle-Noise-Invariant Convolutional Neural Network for SAR Target Recognition

Abstract

Speckle noise is inherent to synthetic aperture radar (SAR) images and degrades the target recognition performance. Deep learning based on convolutional neural networks (CNNs) has been widely applied for SAR target recognition, but the extracted features are still sensitive to speckle noise. In addition, speckle noise has been seldom considered in such CNN-based approaches. In this letter, we propose a speckle-noise-invariant CNN that employs regularization for minimizing feature variations caused by this noise. Before CNN training, we performed SAR image despeckling using the improved Lee sigma filter for feature extraction. Then, we generated SAR images for CNN training by adding speckle noise to the despeckled images. The proposed regularization improves both the feature robustness to speckle noise and SAR target recognition. Experiments on the moving and stationary target acquisition and recognition database demonstrate that the proposed CNN notably improves the classification accuracy compared with the conventional methods.