Computationally Efficient Non-uniform Dehazing Network in Wavelet Domain
Xu Wang, Qingbo Wu · 2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP) · 2022
Data-driven methods for single image dehazing have achieved considerable results in recent years. However, recent methods lead to huge computational and memory requirements, which limits the application of the model in practical scenarios. Observing that the influence of low frequency in wavelet domain on dehazing is much greater than that of high frequency, we propose a non-uniform dehazing network which spends more computational resources on low frequency deblurring and less on high frequency denoising. Our method deblurs the low-frequency components generated by the wavelet transform through a MD (Multi-scale Dehazing) network with self-attention module, and performs denoising operations on high-frequency components through an ED (Enhance-Denoising) network with low computational cost. At the same time color restoration is done by a CAC (Color Adaptive Correction) network seperately. The extensive experiments on synthetic and real datasets show that our method is only slightly inferior to the state-of-the-art methods and outperforms all other methods in dehazing metrics. But our method requires less than half the computation of the best method. So we achieve a good trade-off in performance and computation.