Efficiently Trained Real Image Dehazing Network With Dual Discrete Priors for Enhanced Naturalness

Abstract

In this paper, we efficiently train a dehazing network with enhanced performance by introducing new network architectures and objective functions. Our dehazing network uses high-quality discrete priors from a vector quantization network pretrained on clean images. To mitigate the prolonged pretraining time of existing methods, we analyze the metrics related to discrete priors and propose criteria for early stopping, significantly reducing training time. Furthermore, we introduce dual branches, namely the texture and structure branches, into the dehazing network. The branches act as priors, consisting of pretrained components. To enhance naturalness, we apply our new Structure Alignment Loss with the structure branch which is active only during training, and adopt losses in the frequency domain. Moreover, our analysis of the quantization gap between real and synthetic data shows that additional domain adaptation is unnecessary. Experiments demonstrate that our method outperforms strong baselines on real-world datasets.