<p>Image deblurring aims at recovering clear images from degraded blurred images. From the existing research results, most of learning-based image deblurring networks have a common problem: long training time, large amount of training data, etc. Traditional image deblurring methods can still achieve good deblurring effect without the support of large amount of data, but the computational complexity is relatively high. Aiming at the above problems, this paper proposes a new image deblurring method based on unsupervised network with optimization alternating iterations. Firstly, constructing an unsupervised network based on the encoder-decoder framework, the goal is generating a potentially clear image by inputting random noise. Then, a closed-form solution of the blurring kernel is constructed by combining the cost function of the optimization model. Finally, the loss function is used to optimize the image and the blurring kernel in alternating iterations to generate the final deblurring result. Compared with the state-of-the-art deblurring methods, both objective metrics and visualization experimental results show that the proposed method achieves superior results. The proposed method achieves an average PSNR improvement of 0.32 dB compared to the second-best method. Additionally, the average HNMSE of the blur kernel is reduced by 0.018 compared to the second-best method. Further, the ablation experiments show that the proposed method can effectively combine the advantages of traditional methods and deep learning.</p>