|
今年暑假看了一些Tone-Mapped IQA相关的论文,把记在word上的笔记贴在博客上分享一下吧,欢迎批评指正。
现在Tone-Mapped Image主要有两个数据库: 1、Objective quality assessment of Tone-Mapped images, IEEE Transactions on Image Processing, 2013.
2、No-Reference quality assessment of Tone-Mapped HDR Pictures, IEEE Transactions on Image Processing, 2017.
(这篇论文可以参考一下另外两篇论文:Perceptual quality prediction on authentically distorted images using a bag of features approach /// No reference image quality assessment based on log-derivative statistics of natural scenes)
相关论文包括: Full reference objective quality assessment of Tone-Mapped images ///
IEEE Transactions on Multimedia
Visual attentation guided quality assessment of Tone-mapped images using scene statistics /
FSITM: A feature similarity index for tone-mapped images, IEEE SPL, 2015 ///
Preference of Experimence in image Tone-mapping: dataset and framework for objective measures comparison, IEEE JSTSP, 2017 ///
Blind quality assessment of Tone-mapped images via analysis of information naturalness, and structure.
总结一下:13TIP数据库中的images较少,而且现有方法在这个数据库上的performance还挺高的。17TIP数据库中images的内容形式更多,也更复杂了,有1000多张,三种类型的“distortion”. 相对而言,比较难做。过多的考虑某一种失真的问题,又无法cover另一种失真。好像Deep learning已经刷到挺高的性能了.... 没法活了!!!
=================================================================================================================
1. 2016TMM Blind quality assessment ofTone- Mapped Images via analysis of information, naturalness, and structure
框架:Features + SVR
特征包括三部分:
1)Information: 首先对TM图像进行变换(darkening/brightening the originalluminance),公式为:(Ii=min(max(Mi*I, 0), 255)), M的取值为{1, n, n分之一| n= 3.5,5,6.5,8} ,一共有9种取值。 每一种变换可以得到一种feature map, 计算整幅图像的entropy值 Eg(Global),并将图像进行分成72*72的块,计算得到每一块的entropy的值,再平均,得到El (Local)。
Information = 0.59* Eg + 0.41*El(^1.5)
2) Naturalness: 将图像分成11*11的块,计算块的均值和方差,再计算所有块均值和方差的均值,使用3000张质量完好的图像,提取图像的均值和方差特征,发现符合类高斯模型。分别拟合得到一个函数,得到图像的均值和方差,可以得到两个概率值,取较小者,表示Naturalness。
3)Structure:使用Sobel operator计算梯度图,给定一个阈值,对梯度图进行分割,得到一张二值图,计算该二值图的均值,表示structure。
2. 2017TIPNo Reference quality assessment of Tone-Mapped HDR Pictures
框架是一样的,只是用的features不一样
1)Spatial: Shape and scale parameters of the GGD fitted to the MSCN coefficients(1-2)
2)Spatial: Shape and scale parameters of the GGD fitted to the log-derivative ofthe seven types of neighbors (3-16)
3)Spatial: Two parameters extracted from the σ field (17-18)
4)Gradient: Shape and scale parameters of the GGD fitted to the MSCN coefficientsof gradient magnitude field (19-20)
5)Gradient: Shape and scale parameters of the GGD fitted to the log0derivative ofthe seven types of neighbors of gradient magnitude field (21-34)
6)Gradient: Two parameters extracted from the σ field of gradient magnitude field
7)Gradient: Mean, standard deviation, skewness, and kurtosis of gradientstructure tensor (37-40)
总结:个人觉得论文中给出的features并不能给太多的启发,而是论文中描述的为什么使用这些features, 对我们的帮助更大。后面应该不会研究Tone-Mapped Image Quality Assessment问题了。因为我对Tone mapping也不懂。
|