基于迭代注意力归一化流的低光图像增强

Low-light image enhancement based on iterative attention normalized flow

针对网络层级间特征融合不足并缺乏高频特征的精准定位和获取,以及低光图像和多个正常曝光图像之间的不确定映射问题,提出一种迭代注意力归一化流(Iterative attention normalization flow,IANFlow)网络。迭代注意力模块使用空间和通道注意力对输入特征图的高频特征区域定位后进行特征获取,通过递进式层级定位和融合促使深层特征图包含更多的高频特征;可逆归一化流模块学习低光照图像和正常曝光图像之间复杂的条件分布以及将负对数似然(negative log likelihood,NLL)最小化建立低光图像和参考图像之间一对多的映射。在三个数据集上分别对比LLFlow网络,IANFlow网络的峰值信噪比(peak signal-to-noise ratio,PSNR)分别提高了1.1 dB、1.27 dB、2.14 dB。

An Iterative attention normalization flow(IANFlow)network is proposed to address the problem of insufficient feature fusion between network layers and lack of accurate localization and acquisition of high-frequency features,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 minimizes the negative log-likelihood(NLL)to establish the uncertainty in mappings between a low-light image and a reference 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.