基于知识蒸馏的雷达图像联合降噪超分辨率方法

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

深度学习已在光学图像超分辨率任务中取得成功,但针对合成孔径雷达(Synthetic Aperture Radar,SAR)图像的超分辨率方法仍然有限。这是由于SAR图像固有的散斑噪声和较高的模糊度,使得超分辨率重建更加困难。简单地将去噪和超分辨率模型级联并不能获得最优的重建效果,因为级联方案会使得各模型产生的误差进一步积累。为了在SAR图像超分辨率任务中更好地同时去除噪声并保留纹理细节,本文提出了一种基于知识蒸馏的联合去散斑超分辨率算法。使用干净数据训练教师模型,在训练学生模型时,通过在编码阶段引入特征蒸馏,使得学生网络获得无噪声的潜在变量,从而恢复出干净的高分辨率图像。本文在仿真SAR图像数据集和真实SAR图像数据集上进行了实验。实验结果显示,对于仿真数据集,无论在何种噪声强度下,相比于级联方案,该方法的峰值信噪比(Peak Signal-to-Noise Ratio,PSNR)可以取得约0.4~1 dB的提升,结构相似性指数(Structural Similarity Index Measure,SSIM)可以取得约0.01~0.04的提升。相比于其他方法,该方法的PSNR值取得了约3~9 dB的显著提升,SSIM值取得了约0.01~0.1的提升。对于真实的SAR场景,该方法在等效视数(Equivalent Number of Looks,ENL)、基于平均值比的边缘保留度(Edge Preservation Degree based on the Ratio of Average,EPD-ROA)和比率均值(Mean of Ratio,MoR)等无参考指标上也取得了最佳结果。本文提出的算法在视觉效果上也更有效地抑制了散斑噪声,保留了图像的纹理细节。实验结果验证了本文提出的基于知识蒸馏的SAR图像联合降噪超分辨率方法的有效性,明显优于传统的级联方案和其他算法。

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.