ISAR image enhancement based on a Recursive Residual Network

Applying convolutional neural networks(CNNs) to inverse synthetic aperture radar (ISAR) image enhancement has proven effective. Compared with traditional sparsity-driven methods, recently developed ISAR image enhancement methods based on CNNs offer significant improvements in resolution enhancement, sidelobe suppression, and imaging efficiency. However, these methods may show outstanding performance in the same scenarios where the training data was generated and exhibit suboptimal performance when the scenario changes. Additionally, all these methods adopted the mean-squared error as the loss function, causing the reconstructed ISAR image to lose high-frequency information and the texture is too smooth. To address these limitations, a modified ISAR image enhancement framework based on a recursive residual network is proposed. The ISAR image enhancement processing framework and the training dataset generation method were improved to minimize the influence of the fixed imaging scenarios. A combined loss function, composed of the multi-scale structural similarity loss and the smooth L1 loss, was adopted to retain the high-frequency information and the edge details of the ISAR image, while improving the resolution. Experimental results demonstrated that compared with extant methods, the proposed method provides reconstructed ISAR images with higher resolution and improved visual quality over a range of different scenarios.