Real-aware motion deblurring using multi-attention CycleGAN with contrastive guidance

Most existing motion deblurring methods need a large amount of paired sharp and blurred images for network training. However, this restricts network representation ability and fails to maintain satisfactory deblurring results under real-world circumstances. To overcome these limitations, we propose an unsupervised real-aware motion deblurring method using multi-attention CycleGAN with contrastive guidance. The network architecture has two streams, including a forward sharp translation stream and a backward blurred translation stream, to handle unpaired sharp and blurred images based on a cycle-consistent mechanism. First, we develop a multi-attention GAN for each translation stream to embrace real-aware information from unpaired real sharp and blurred statistics. The multi-attention includes a long-short attention block, a multi-kernel attention block, and an adversarial attention block. Second, we propose a gradient contrastive loss function in the generator and an adversarial contrastive loss function in the discriminator. They exploit inherent sharp information and increase network representation in practical applications. Third, we design hybrid loss functions for sharp and blurred translations to train the network. Extensive experiments on four benchmark datasets demonstrate that the proposed method achieves better-restored performance than current state-of-the-art methods for unsupervised motion deblurring.