Global Cognition and Local Perception Network for Blind Image Deblurring

Chuanfa Zhang, Wei Zhang, Feiyu Chen, Yiting Cheng, Shuyong Gao, Wenqiang Zhang

Abstract: Nowadays, people are more and more used to taking pictures with handheld devices. However, the photos taken by handheld devices are easy to blur. Recent state-of-the-art methods for image deblurring often restore image through multiple stages, which leads to excessive network parameters and slow running speed, and makes it difficult to apply to mobile devices. While one-stage image deblurring methods have fewer parameters and faster running speed, the performance of their models is not so satisfactory. To solve these problems, we propose a lightweight one-stage image deblurring network called Global Cognition and Local Perception Network (GCLPN) in this paper. We design the Global Cognition Module (GCM) to obtain global feature descriptors and present the Local Perception Module (LPM) to help image reconstruction in local regions. Furthermore, we introduce a gradient loss that focuses on subtle image texture information during network training. Experimental results illustrate that our GCLPN surpasses the state-of-the-art methods in standard metrics of PSNR and SSIM with the least number of parameters (about 1.7M) and real-time running speed (about 0.37 s per $$720 \times 1280$$ image).