Single ISAR image enhancement based on convolutional neural network

Abstract

Compared with traditional sparsity-driven methods, inverse synthetic aperture radar (ISAR) image enhancement method based on convolutional neural networks (CNNs) have outstanding performance in recent research, which improved the resolution of reconstructed image significantly with higher imaging efficiency. However, recently developed ISAR image enhancement methods based on neural networks are only effective in the same scenarios where the training data was generated. Additionally, all these method adopted the mean-squared error as the loss function, causing the reconstructed ISAR image to lose high-frequency information and fail to capture appropriate details. To address these limitations, a single ISAR image enhancement framework based on a modified super-resolution convolutional neural network (SRCNN) is proposed in this paper. The ISAR image enhancement processing framework were improved to minimize the influence of the fixed imaging model. A combined loss function, composed of the structural similarity (SSIM) loss and the L1 loss functions, was adopted in the proposed framework to retain the high-frequency information and the luminance information of the ISAR image, while improving the resolution. Through quantitative analysis of experimental results by using different quality evaluation indicators, it demonstrated that compared with extant methods, the proposed framework provides reconstructed ISAR images with higher resolution and definition over a range of different scenarios.