無參考圖像質量評價(PIQE)算法:Blind Image Quality Evaluation Using Perception Based Features

paper:Blind Image Quality Evaluation Using Perception Based Features
date:2015
authors:Venkatanath N* etc...
code:PIQE

1.Introduction

Firstly, humans visual attention is highly directed towards salient points in an image or spatially active regions. Secondly, local quality at block/patch level adds up to the overall quality of an image as perceived by humans. In our approach, the first principle is addressed by estimating distortion only on regions of spatial prominence and the second, by computing distortion levels at the local block level of size n × n, where n = 16. Working at the block level would enable us to exploit the local characteristics that account for overall perceptual quality of an image.

1.主觀質量評價對圖像中某些顯著重要的部分更為關注。2.由局部塊的質量分數得到整體質量分數。
作者提出了方法Perception-based Image Quality Evaluator(PIQE)

2.PIQE

A.Divisive Normalization

類似BRISQUE，先計算MSCN系數\(\hat{I}(i,j)\)。
將\(\hat{I}(i,j)\)分成\(N_b\)個size為\(n \times n\)的block(n=16)，利用\(\hat{I}(i,j)\)標記每一個block是uniform(U)還是spatially active(SA)。

\[B_k=\left \{\begin{aligned} U && \nu_k<T_U \\ SA & & \nu_k \ge T_U \end{aligned} \right. \]

其中\(\nu_k\)是塊\(B_k\)　MSCN系數的方差，\(k\in 1,2,\dots,N_b\), \(T_U\)是預設的值，作者設為0.1。下面只對SA塊評估質量分數。

B.Block Level Distortion Estimation

失真可分為三類:塊效應(blockiness)，模糊(blur)，噪聲(noise)。
作者對spatially active \(B_k\)失真分兩類處理，noticeable distortion和white noise.

1.Noticeable Distortion Criterion

對於\(n \times nB_k\)的每一條\(L_p\),划分為11個segment

\[a_{pq}=L_p(x):x=q,q+1,\dots,q+5 \]

\(p\in 1,2,3,4\)表示四條邊，\(q\in 1,2,3,\dots,11\)表示每條邊可分割11個segment，每一segment長為5。
如果有任何一個segment標准差小於某個閾值，則是low spatial activity,視為存在noticeable distortion，即滿足

\[\sigma_{pq}<T_{STD}　\tag{1} \]

2.Noise Criterion

將block划分為中心區域和周邊區域，分別計算標准差為\(\sigma_{cen}\)和\(\sigma_{sur}\),計算block的標准差為\(\sigma_{blk}\)

\[\beta = \frac{\vert\frac{\sigma_{cen}}{\sigma_{sur}}-\sigma_{blk}\vert}{\max(\frac{\sigma_{cen}}{\sigma_{sur}},\sigma_{blk})} \]

如果存在noise，則有如下關系:

\[\sigma_{blk} > 2*\beta \tag{2} \]

3.Quantifying Distortion using Block Variance Parameter

上面提出兩個准則(1),(2)判斷是否存在兩種失真，然后作者使用block的variance來度量兩種失真。

we could observe that the variance of the MSCN coefficients of a given block, \(\nu_{blk}\) shows significant signature of the amount of distortion present in that block.

4.Pooling

使用\(v_{blk}\)度量block的失真，

\[D_{sk}=\left \{\begin{aligned} 1 && if& (1)and(2) \\ \nu_{blk} && if& (2)\\(1-\nu_{blk}) & & if& (1) \end{aligned} \right. \]

整張圖片的質量分數:

\[PIQE=\frac{\sum_{k=1}^{N_{SA}}D_{sk}+C_1}{N_{SA}+C_1} \]

其中\(C_1\)是為了數值穩定性的常數，\(N_{SA}\)是spatially active blocks的總數。

3.Experiments