Improved Image Dehazing Algorithm Based on Multi-scale Feature Fusion
Xinyu Liu, Jie Yang, Wusheng Shang · 2021 IEEE 6th International Conference on Signal and Image Processing (ICSIP) · 2021
To address the issue of patch effect and halo effect in dehazed images due to the influence of dehazing, an improved image dehazing algorithm is proposed based on multi-scale feature fusion. On the one hand, the patch effect and halo effect expose the inaccurate estimation of parameters in image dehazing, the transferring efficiency of image features is enhanced by introducing an attention mechanism in this paper, which shows a more accurate estimation and better convergence ability. On the other hand, for a more real dehazing effect, the realness index is introduced to the loss function in the network. Compared with the recent state-of-the-art dehazing methods, the results of the validation in the NYU-depth2 test set show the peak signal-to-noise ratio index of improved image dehazing network is increased by 7.28%, and the structural similarity index reaches 0.901. The improved image dehazing network can effectively restore images from haze and the dehazed images are more real and natural.