Joint Denoising and Super-Resolution Method for Radar Images Based on Knowledge Distillation

‍ ‍Deep learning has significantly improved the task of super-resolution （SR） for optical images. However， SR methods for synthetic aperture radar （SAR） images remain limited owing to the inherent speckle noise and high ambiguity in SAR images， thus rendering SR reconstruction challenging. Cascading denoising and SR models does not yield optimal reconstruction results as the cascading approach can result in the accumulation of errors generated by each model. To simultaneously remove noise and preserve texture details in the SR task for SAR images more effectively， this paper proposes a joint despeckling and SR algorithm based on knowledge distillation. First， a teacher model is trained using clean data. When training the student model， feature distillation is introduced during the encoding stage， which allows the student network to acquire noise-free latent variables， thereby allowing the recovery of clean high-resolution images. Experiments are conducted on simulated and real SAR image datasets. The results show that for the simulated datasets， the proposed method improves the peak signal-to-noise ratio （PSNR） and structural similarity index measure （SSIM） by 0.4~1 dB and 0.01~0.04， respectively， compared with the cascading approach， regardless of the noise intensity. Furthermore， the method significantly improves the PSNR and SSIM by 3~9 dB and 0.01~0.1， respectively， compared with other methods. For real SAR scenes， the proposed method achieves the best results for no-reference metrics such as the equivalent number of looks， edge-preservation degree based on the ratio of average， and mean of ratio. Visually， the proposed algorithm suppresses speckle noise more effectively while preserving image texture details compared with other approaches. The experimental results validate the effectiveness of the proposed knowledge-distillation-based joint despeckling and SR method for SAR images， as evidenced by its performance， which significantly surpasses those of conventional cascading approaches and other algorithms.