Low-light image enhancement based on iterative attention normalized flow
An Iterative attention normalization flow(IANFlow)network is proposed to address the problem of in-sufficient feature fusion between network layers and lack of accurate localization and acquisition of high-frequency fea-tures,as well as the problem of uncertain mapping between low-light images and multiple normal-exposure images.The iterative attention module uses spatial and channel attention to localize the high-frequency feature regions of the input feature maps and then performs feature acquisition,which prompts the deeper feature maps to contain more high-frequency features through incremental hierarchical localization and fusion;the reversible normalization flow module learns the complex conditional distributions between low-light images and normal-exposure images as well as minimi-zes the negative log-likelihood(NLL)to establish the uncertainty in mappings between a low-light image and a refer-ence image.one-to-many mapping.The peak signal-to-noise ratio(PSNR)of the IANFlow network is improved by 1.1 dB,1.27 dB,and 2.14 dB when comparing the LLFlow network on each of the three datasets.
NETL
NSTL
万方数据
