基于共轭梯度法的全变差盲图像去模糊仿真

The Image Visual Effect is Exactly The Same as the Original Test Image Structure Information

为解决当前方法盲图像去模糊视觉效果较差、原始图像细节特征丢失的问题,提出了基于共轭梯度法的全变差盲图像去模糊方法。该方法利用shock滤波器从模糊目标图像中预测清晰边缘梯度和模糊边缘梯度,模拟目标图像质量退化过程,将预测得到的清晰边缘梯度作为先验知识,采用尺度策略实现运动模糊核的估计;采用共轭梯度法将目标图像的共轭性与图像已知像素点的梯度值构成一组共轭方向并沿着该方向进行全局搜索,实现误差代价函数的极小值迭代求解;在此基础上,将二阶差分最小化约束条件加入到求解目标盲图像误差代价函数极小值中去,采用全变差正则化方法,将目标盲图像去模糊问题转换为频域滤波问题处理。仿真结果表明,提出方法有效提高了去模糊后图像的峰值信噪比,视觉效果与原始测试图像结构信息基本一致,同时提升了图像的去模糊效率。

This article puts forward a deblurring method for total variation blind image based on conjugate gradient method. This method used the shock filter to predict the clear edge gradient and the fuzzy edge gradient from the fuzzy target image and simulated the degradation process of target image quality. Then, the method regarded the predicted clear edge gradient as the prior knowledge and used the scale strategy to estimate the motion blurring kernel. In addition ,our method used the conjugate gradient method to combine the conjugation of target image with the gradient values of known pixels in image, and thus to form a set of conjugate directions, and then we searched globally along this direction, so as to realize the iterative solution of minimum value of error cost function. On this basis, we added the second-order difference minimization constraint to the solution for the minimum value of error cost function of target blind image. Finally, the total variation regularization method was adopted, so that the problem of deblurring the target blind image was transformed into the problem about frequency domain filtering. Simulation results show that the proposed method effectively improves the peak signal-to-noise ratio of the image after deblurring. The visual effect is basically consistent with the structure information of original test image. Meanwhile, the deblurring efficiency of image is improved.