无监督学习全局和局部特征建模的低光照图像增强

Low-light Image Enhancement Based on Unsupervised Learning Global and Local Feature Modeling

在EnlightenGAN的启发下,提出了一种新的基于无监督学习全局和局部特征建模的低光照图像增强网络(Low-light Image Enhancement Network Based on Unsupervised Learning Global and Local Feature Modeling Image Enhancement,GLFMIE)。该网络分为两个阶段:生成网络和判别网络。生成网络包括全局和局部特征建模网络,判别网络包括全局和局部判别网络。在全局特征建模中创新性地引入了Swin-Transformer Block,其移位窗口机制可以以较少的内存消耗对输入图像进行长距离的特征依赖建模,并很好地提取图像颜色、纹理和形状的特征,从而有效地抑制噪声和伪影。在局部特征建模中,设计了一种多尺度图像和特征聚合(Multi-Scale Image and Feature Aggregation,MSIFA)网络,允许在单个U型网内交换来自不同尺度的信息,进一步增强图像特征的表征能力。在多个公共数据集的测试实验中,与已有一些先进低光照图像增强算法相比,该算法均取得了SOTA级别的表现。

Inspired by EnlightenGAN,a new Low-light Image Enhancement Network Based on Unsupervised Learning Global and Local Feature Modeling Image Enhancement(GLFMIE)is proposed.The network is divided into two stages:generator network and dis-criminator network.The generator network includes global and local feature modeling network,and the discriminator network includes global and local discriminator network.Swin-Transformer Block is introduced in global feature modeling.Its shift window mechanism can model a long distance feature dependence of the input image with less memory consumption,and extract the features of image color,tex-ture and shape well,thus effectively suppressing noise and artifacts.In local feature modeling,a Multi-Scale Image and Feature Aggrega-tion(MSIFA)network is designed to allow exchange of information from different scales in a single U-shaped network,further enhan-cing the representation ability of image features.In test experiments of multiple public datasets,compared with some advanced low-light image enhancement algorithms,this algorithm has achieved SOTA level performance.