Data-Free Learning for Lightweight Multi-Weather Image Restoration

Image restoration has made a remarkable performance with the large-scale training data and increasing model capacity. However, the burdensome model complexity hinders the mode deployment on resource-constrained devices. Besides, the training data may be unavailable due to some constraints, which undoubtedly affects the efficient model learning. In this paper, we propose an effective data-free model compression framework for lightweight multi-weather image restoration, which consists of data generation and model distillation stages. Specifically, a data generator is first utilized to synthesize degradation-aware samples from a latent distribution. Then, the on-the-shelf teacher model provides a pseudo-label to supervise the training of the student model. To ensure the diversity of the training data, adversarial learning is adopted to maximize the dependency between teacher and student models. Moreover, we adopt a contrastive regularization constraint to further improve model representation. Experimental results show that our proposal achieves comparable performance with the student model trained with the original data and some unsupervised methods for image dehazing and deraining tasks.