摘要
由于现实环境中明暗光照的剧烈变化,现有的低光照图像增强方法往往会导致增强后的图像亮度和对比度不足,出现伪影和模糊等情况。此外,当前的低光照图像增强工作仅针对于图像亮度的提升,而对于噪声影响的处理较少,这些都不利于低光图像的增强。为了解决上述问题,论文提出了一种基于上下文Transformer的低光照图像增强算法。具体地,论文首先利用动态卷积网络对低光照图像进行特征提取;接着,设计了上下文Transformer对得到的特征图进行全局关联的深层特征提取,并使用金字塔池化模块进行去噪处理;最后,通过瓶颈结构的卷积网络输出得到增强后的图像。在多个主流数据集(LOL,LIME,DICM等)上的对比实验结果表明,与目前已有的主流工作相比,论文所提方法的结果不仅在主观视觉上有更好的视觉效果,更加符合人眼的视觉特点;而且在各种定量客观评价指标上也有良好的表现,尤其在PSNR和SSIM两个指标上有明显的提升。
Due to the dramatic changes of light and dark in the real world,existing methods of low-light image enhancement often leads to insufficient brightness and contrast of the enhanced image,resulting in artifacts and blurring.In addition,the existing low-light image enhancement work is only aimed at improving the brightness of the image,and less processing of the influence of noise,which are harmful to the low-light image enhancement.To solve these problems,this paper proposes a low-light image en⁃hancement algorithm based on contextual transformer.Specifically,first of all,this paper uses dynamic convolution network to ex⁃tract features from low-light images.Then,the context transformer is designed to extract the deep features of the obtained feature map by global correlation,and use the pyramid pooling module to denoise.In the end,the enhanced image is obtained through the bottleneck structure convolution network.The comparison results on several mainstream datasets(LOL,LIME,DICM,etc.)show that compared with the existing mainstream work,the results of the method proposed in this paper not only have better visual effects in subjective vision,but also have better visual effects,it conforms to the visual characteristics of the human eye,and it has a good performance in various quantitative indicators,especially in PSNR and SSIM.
作者
徐文晨
樊佳庆
宋慧慧
XU Wenchen;FAN Jiaqing;SONG Huihui(School of Automation,Nanjing University of Information Science and Technology,Nanjing 210044;School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106)
出处
《计算机与数字工程》
2023年第1期237-244,共8页
Computer & Digital Engineering
基金
国家自然科学基金项目“基于深度学习的多源多分辨率遥感影像融合算法研究”(编号:61872189)资助
