Contrastive Learning Aided Single Image Deblurring

Fast and high-quality image deblurring is essential for real-time applications on edge devices. Reduction of the solution space is important for the ill-posed deblurring task. Different from most existing deblurring networks that pull together a blurry image and a sharp image in the latent space, this paper proposes contrastive learning aided deblurring that also pushes a deblurred image apart from a blurry image for training the existing MIMO-UNet. The contrastive loss combined with the multi-scale content loss and frequency reconstruction loss helps the deblurring network pull together the anchor and positive pair while pushes the anchor apart from the negative pair. The proposed progressive negative sample scheme gradually increases the lower bound of the quality of negative samples to keep the contrastive loss decreasing. For the GoPro dataset, the proposed scheme improves the quality of MIMO-UNet while keep the low inference time (0.392 second on GPU 1080Ti) and number of parameters (6.8 M) of the network unchanged.