基于稀疏正则化和深度学习的湍流退化图像复原研究

Aero-optics effects become a critical problem in precise guidance and pinpoint strike. It affects accuracy of long range imaging system, and even the development of precision strike weapons. However, the captured images through a turbulent medium usually simultaneously exhibit highly random geometric distortion and space-time-varying blur. Hence, we propose a restoration method of turbulence-degraded images, including distortion correction, deblurring, and single image super-resolution. First, the proposed scheme constructs a high-quality reference image from the observed frames via low-rank decomposition, and the reference image is iteratively optimized through sparse regularization. After registering the observed frames to the optimized reference image, the non-rigid geometric distortion is well corrected. Second, by using locally adapted data-driven kernel estimation, the proposed method can fuse the registered frames into single image with reduced blur variation, transferring the space and time-varying problem to a shift invariant one. Applying a blind deconvolution algorithm the fused image will remove the turbulent blur. Next, the proposed scheme further improves the quality of the restored image by a deep learning based super-resolution reconstruction method, which can capture high-order dependencies of texture information by using hierarchical models that extract highly non-linear representations of the massive amounts of high-resolution images. Finally, we conduct experiments on the simulated and real turbulence-degraded images to demonstrate the effectiveness and robustness of our proposed method.

气动光学问题已经成为精确制导和精确打击的关键，该研究的进展将直接影响高速飞行器光学探测系统的精确性，以及新一代精确打击武器的发展。在气动湍流环境下，高速飞行器远距离目标检测的公认难题是解决成像的强畸变、高度随机的严重模糊问题，即湍流退化效应。针对正在承担的气动光学效应校正、目标识别工程任务，我们提出一种畸变图像校正、模糊图像复原、复原图像的超分辨重建的理论和方法。首先，针对湍流退化序列图像存在严重的畸变问题，提出低秩矩阵复原和稀疏正则化约束的参考图像构造与优化算法，实现高精度的湍流畸变校正；其次，为了降低时空模糊的随机性，提出基于核估计理论的图像融合方法，对融合结果进行盲去卷积，实现快速湍流模糊复原；然后，研究构造深度神经网络，挖掘海量高分辨图像的层次化高阶纹理先验，实现湍流复原图像结构和纹理的超分辨率增强；最后，结合承担工程任务的仿真与真实数据集，验证本研究成果的鲁棒性、有效性。

气动光学问题已经成为精确制导和精确打击的关键，该研究的进展将直接影响高速飞行器光学探测系统的精确性，以及新一代精确打击武器的发展。在气动湍流环境下，高速飞行器远距离目标检测的公认难题是解决成像的强畸变、高度随机的严重模糊问题，即湍流退化效应。本课题实现了基于稀疏正则化和深度学习的序列多帧湍流退化图像校正复原的三个目标。首先，在理论方法创新方面，提出了求解稀疏正则化约束的快速Bregman迭代复原算法并证明了其收敛性，同时还提出了基于加权分数核范数的新型低秩矩阵复原理论，进一步完善湍流退化图像复原的理论基础。其次，在关键技术突破方面，提出了基于稀疏正则化约束的多帧湍流图像畸变校正和基于形变场引导的时-空核回归多帧融合方法；同时，为了降低传感器噪声的干扰，提出了基于加权分数核范数的高斯噪声以及混合噪声去除方法；为了复原退化图像的局部细节，提出了基于稀疏表示和多核学习的超分辨率重建算法，完成了课题对湍流退化图像复原的关键技术研究目标的要求。最后，在验证性应用方面，本课题的理论方法和关键技术均应用于国家某重大战略工程，成功服务于国防科学基础建设，并且在结题考核中获得了极高的评价。