提出了一种基于L0稀疏先验的改进正则化模糊图像盲复原算法来解决相机抖动所产生的模糊问题。根据模糊图像的梯度分布要比清晰图像稠密并且暗通道的稀疏性也相对较小这一固有属性建立了新的优化模型。针对L0范数的高度非凸性和暗通道稀疏优化过程中涉及到的非线性最小化问题, 提出了一种近似线性映射矩阵, 并用半二次分解法对L0最小化问题进行求解。最后, 采用快速傅里叶变换在频域中对模糊核及清晰图像进行交替迭代运算得到复原图像。对多幅不同类型的模糊图像进行了实验, 结果显示: 复原图像平均灰度梯度高达11.411, 图像信息熵达到7.304, 处理365×285的图像只需8.07 s。提出的算法有效抑制了图像边缘处的振铃效应, 完整保留了清晰的细节信息的同时显著提高了运算速度, 并适用于多种不同类型图像的盲复原。

An improved regularization blind restoration method based on L0 sparse prior was proposed to overcome the image blue from camera shake. A new optimization mode on the basis of inherent property which the gradient distribution of the blurred image is denser than that of the clear image and the sparse of the dark channel is relatively smaller. Aiming at the highly non-convex of L0 norm and nonlinear minimization problem in the dark channel sparse optimization process, an approximate linear map matrix based on look-up tables was proposed, and the linearized L0 minimization problem was solved by half-quadratic splitting methods. Finally, the fast Fourier transform was used to do iterative operation alternately for the fuzzy kernel and the clear image in frequency domain to obtain the restored image. Through experiments on several different types of blurred images, the results show that average gray level gradient is up to 11.411, the image entropy is up to 7.304, and it only takes 807s to process 365×285 images. The improved regularization algorithm effectively suppresses the ringing effect near the edge of the image, retains the integrity of clear details, improves the speed of operation significantly. The algorithm is suitable for all kinds of image restoration.