GAN-Based Encoder-Decoder Method for Image Deblurring

Image deblurring is one of the classic tasks in the field of computer vision, holding significant research importance and practical value. The Multi-Input Multi-Output (MIMO) encoder-decoder network, characterized by its small parameter size and low training costs, has been adopted for image deblurring. However, this method still suffers from poor generation quality and insufficient capability to capture details and textures. To address these issues, this paper proposes a refined network structure based on MIMO-UNet to improve image quality. The network is optimized across three principal dimensions. Firstly, it integrates a generative adversarial network (GAN) where the generator is structured as a MIMO-UNet and introduces a discriminator, forming a GAN. Secondly, within this GAN framework, the traditional residual blocks in the generator are substituted with Frequency Feature Extraction (FFE) Blocks, which augment the utilization of frequency domain features. Thirdly, the generator's total loss includes additional frequency domain L2 loss, adversarial loss, and edge loss, which improves the model’s ability to extract key features. Experimental results demonstrate that compared to the baseline MIMO-UNet, the proposed method has increased the PSNR metric by 3.54% and the SSIM metric by 0.63%. Overall, the results are satisfactory.