DefoGAN: an end-to-end single defocus image deblurring method

Defocus blur is one of the most commonly arising types of artifacts when taking photos. Blind defocus image deblurring is a classical ill-posed problem, as we have to estimate two variables with one observation. The blur in the defocus image is usually spatially varying which makes the estimation of clear image more challenging. In this paper, we propose an end-to-end blind defocus deblurring network (DefoGAN) by repeatedly incorporating an De-RES Block. DefoGAN network is based on the generate adversarial network, while utilizing feature pyramid network (FPN) for multi-scale feature extractor. The proposed DefoGAN is shown to improve the generated images’ quality and authenticity and achieve a balance of quality and speed than other compared methods. Ablation studies and evaluation metrics proved the effectiveness of the proposed method.