基于维纳滤波器和生成对抗网络的动态模糊图像处理方法

Dynamic Fuzzy Image Processing Based on Wiener Filter and Generative Adversarial Networks

在傅里叶频域中,由于逆滤波对加性噪声特别敏感,使得恢复后的图像仍然非常模糊.针对这一问题,我们提出了一种基于维纳滤波器和生成对抗网络的动态模糊图像处理方法.首先使用维纳滤波去模糊算法,通过均方差最小化去除噪声,但由于无法判断拍摄装置的移动范围并未得到预期效果.再考虑使用自由性强、不受预定条件分布的生成对抗网络模型(GAN).定义一个类生成器G(y)和类判别器D(x),通过机器学习的方式进行反复学习和反馈,直至达到模型无法判别生成数据样本S(y)和真实数据样本r(x)时,图像近似还原成功.同时,引入"模糊核"概念,模拟图像的模糊轨迹,进行精确还原.最后,由于肉眼很难对图像的还原程度做定量判断.因此我们利用三个评价指标对这些图像进行客观评价——峰值性噪比PSNR、模糊系数KBlur、质量因素Q.实验结果表明,在该方法下的图像的三个评价指标在一定程度上有所改善,从而得到图像还原较为成功的结论.

In the Fourier frequency domain,the restored image is still very fuzzy because the inverse filter is particularly sensitive to additive noise.In view of this problem,we propose a dynamic fuzzy image processing method based on Wiener filter and Generative Adversarial Networks(GAN).Firstly,the Wiener filter is used to blur the algorithm,and the noise is minimized by mean variance minimization.But expected result is not achieved because it is impossible to judge the moving range of the shooting device.Then,we consider to use a free unpredetermined distribution of GAN model,define a class generator G(y)and a class discriminant D(x),with machine learning approach to learn and feedback repeatedly,until the generated model cannot distinguish the data sample S(y)and real data samples r(x),the image is approximately restored successfully.At the same time,the concept of'fuzzy kernel'is introduced to simulate the fuzzy trajectory of the image and to make precise restoration.Finally,it is difficult to make quantitative judgment on the restoration degree of the image.Therefore,three evaluation indexes are used to objectively evaluate these images—Peak Signal-to-Noise Ratio(PSNR),fuzzy coefficient KBlur,and quality factor Q.The experimental results show that the three evaluation indexes of the image under this method are improved to some extent,so as to obtain a more successful conclusion of image restoration.