基于深度学习的复杂退化模糊图像恢复

Real blurred images usually degraded with certain nonlinear effects, such as image lossy compression and image noises. Such nonlinear factors could reduce the quality of the blurring kernel estimation and also make the traditional deblurring model less reliable. To effectively improve the deblurring quality, we investigate the deep network to estimate the convolution kernel followed by latent image estimation. There are two contributions in our work. The first is the kernel estimation based on Deep Believe Network. We employ the Denoising Autoencoder to pretrain the network thus the robustness of the kernel estimation is improved. The second is the sharp images' Priors learning based on Restricted Boltzmann Machine. We use unsupervised learning to fit the image patch distribution. Then the learned Priors are effectively incorported into the deblurring model. The program would make contribtions in improving the robustness of the image deblurring system.

实际模糊图像通常受多类退化因素共同影响，例如图像压缩失真和图像噪声，这些退化因素会使传统卷积模型无法准确估计卷积核并降低图像恢复品质。本项目研究基于深度学习的方法估计卷积核，再恢复出清晰图像。该项目创新点包括：基于深度信念网络和降噪自动编码机的深度学习方法估计卷积核，抑制噪声和非线性退化因素对卷积核估计的影响。其次，研究基于受限玻尔兹曼机的清晰图像先验特征学习，用无监督学习方法获得先验特征，并将其用于约束反卷积算法。本项目在提高实际模糊图像卷积核估计和去模糊算法的鲁棒性方面做出贡献。

深度学习技术最近几年发展迅速，在计算机视觉、自然语言处理、医学图像处理等各个领域都有获得了重要进展。该项目基于深度学习和传统图像处理方法研究了图像去模糊、图像增强以及图像超分辨率等基础图像处理算法。在传统图像处理模型基础上引入卷积神经网络进一步优化模型，使得模型输出的图像品质更高。项目研究过程中提出的算法已经成功发表在多个SCI期刊以及CCF A类国际会议。