基于像素级和块级的低照度图像增强-《计算机技术与发展》

文章信息/Info

Title:: A Low-light Image Enhancement Algorithm Based on Pixel Level and Block Level

Author(s):: WANG Shi-ju; WANG Xin; JU Ming-ye*; School of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China

关键词:: A Low-light Image Enhancement Algorithm Based on Pixel Level and Block Level

Keywords:: image enhancement; HSV color space; gamma correction; atmospheric scattering model; guided filter

摘要:: 针对低照度图像对比度低、亮度弱、色彩暗淡等问题,提出一种基于像素级和块级的低照度图像增强算法。该算法在 HSV 空间对亮度通道和饱和度通道分别进行像素级、块级增强。前者通过伽马校正构造一种新的像素级增强模型,其采用增强矩阵代替单一伽马值,并结合大气散射模型与全局搜索策略求得模型中的未知参数,进而对亮度通道进行像素级增强;后者着重关注色彩饱和度的提升,将饱和度通道分为若干块,假设每一块具有相同的增强因子,利用约束信息对每个块采用局部一维搜索策略确定其值。将处理后的各通道分量转化至 RGB 空间,获得最终增强结果。该算法有效结合了像素级处理的低复杂度和块级处理的信息丰富度等优势,且不需要任何的训练过程。实验结果表明,在合成数据集与真实场景下,所提算法对亮度的提升和色彩的恢复均有明显改善,在客观评价指标上同样取得优异性能。

Abstract:: Images captured under poor illumination or at night time have weak brightness and dim color. To resolve this,a low - lightimage enhancement algorithm based on pixel level and block level is proposed, which enhances the brightness channel and saturationchannel at pixel level and block level respectively in HSV space. The former constructs a new brightness channel pixel level enhancementmodel via gamma correction,which uses the enhancement matrix to substitute a single gamma value. Furthermore,a method combiningan atmospheric scattering model and global search strategy is proposed to solve the parameters. The latter divides the saturation channelinto several blocks to improve the color saturation. Specifically,it assumes that each block has the same enhancement factors,which areobtained by using the one-dimensional local retrieval strategy of constraint information. The processed channels are converted into RGBspace to achieve the end result. The proposed algorithm effectively combines the benefits of low complexity of pixel level processing andinformation richness of block level processing,and does not require any training process. Experiments show that the proposed algorithmcan dramatically improve the brightness and color saturation in synthetic datasets and real scenes,and also achieve excellent performancein objective metrics.