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.