频域特征蒸馏的尺度融合图像去雾网络

打开文本图片集

中图分类号：TP391. 4 文献标识码：Adoi：10. 37188/OPE. 20253306. 0916 CSTR：32169. 14. OPE. 20253306. 0916

Dual scale fusion image dehazing algorithm based on frequency-domain feature distillation

CHEN Qingjiang，YANG Shuang

（Xi'an University of Architecture and Technology，School of Science，Xi'an ，China） * Corresponding author， E -mail：ysdt917@163. com

Abstract：Aimed at the issue that the edge details of the dehazing images were insufficiently clear，and the majority of the existing U-Net dehazing networks did not adequately exploit the information in the frequen⁃ cy domain and neglected the information exchange among different channels，resulting in a blurry struc⁃ ture，a dual-scale fusion network with frequency-domain feature distillation was proposed for the effective dehazing of single images. In the Coarse-scale feature extraction subnet，a large-scale convolution kernel was utilized to extract image texture information，and a residual attention mechanism was employed to en⁃ hance the features related to haze. In the Fine-scale high-frequency fusion subnet，a high-frequency feature distillation module was devised to refine the extracted structure and edge information and gradually restore clear images. Meanwhile，the cross-fusion strategy was adopted to fuse the features of different channels. The experimental results indicate that compared with the MSTN algorithm（Efficient and Accurate MultiScale Topological Network），the peak signal-to-noise ratio and structural similarity on the outdoor image dataset have been enhanced by 9.98% and 4.77% respectively. The experimental results on diverse datas⁃ ets demonstrate that the proposed approach exhibits superior performance. This method can effectively en⁃ hance the dehazing effect，retain more structural information，and possess better color detail recovery capa⁃ bility.

Key words：high-frequency information；characteristic distillation；image dehazing；cross fusion；residu⁃ al attention

1 引 言

在雾霾和多尘等天气条件下观测或捕获的图像往往会出现对比度低、能见度严重下降的问题。（剩余15907字）