ITS Video Quality Measure

ITS Model [138,142,143,151] maps image imperfections onto measurable mathematical parameters. It is based only on the luminance values, the chrominance values are not considered. From that, we can see that the correlation with subjective tests cannot be very good, as the chrominance values are important to the HVS, even if its importance is not as high as that of luminance values. The method is based on the extraction and classification of features from both the original and distorted video sequences. The extracted features quantify the impairments of the distorted sequence. An algorithm that segments each video frame into motion and still parts is used to quantify each of them separately. The features correspond to the impairments of the sequence. Each feature is evaluated objectively. The classification part is used to assign a quality measure based on the features objective values. The quality measure is a linear combination of three quality impairment measures. These measures were selected among a number of candidates such that their combination matched best the subjective evaluations. One of these measures quantifies the Spatial Information (SI), while the other two measures quantify the Temporal Information (TI). The SI for a frame $F_n$ is defined as

$$SI(F_n) = STD_{space}\{Sobel[F_n]\},$$

where $STD_{space}$ is the standard deviation operator over the horizontal and vertical spatial dimensions (the $space$) in a frame, and $Sobel$ is the Sobel filtering operator, which is a high pass filter used for edge detection. The TI is based upon the motion difference between two successive frames, $\triangle F_n=F_n-F_{n-1}$, which is composed of the differences between pixel values at the same location. TI is computed using

$$TI(F_n)=STD_{space} \{\triangle F_n\}.$$

Both SI and TI are to be computed for each frame. To obtain a single scalar quality estimate for each video sequence, SI and TI values are then time-collapsed as follows. The three measures, $m_1$, $m_2$, and $m_3$, are defined, which are to be linearly combined to get the final quality measure. Measure $m_1$ is a measure of spatial distortion, and is obtained from the SI features of the original and degraded video as follows:

$$m_1=RMS_{time} \left\{5.81\left\vert\frac{SI(O_n)-SI(D_n)}{SI(O_n)}\right\vert\right\},$$

where $O_n$ and $D_n$ are the frame $n$ of the original and degraded video sequences respectively. $RMS_{time}$ denotes the root mean square function performed over the time, for the duration of each test sequence. $m_2$ and $m_3$ are measures of temporal distortion and are given by

$$m_2=f_{time} \left(0.108 \max{\left\{TI(O_n)-TI(D_n),0\right\}}\right),$$

with

$$f_{time}(x)={\textstyle STD_{time}}\{CONV(x,[-1,2,-1])\},$$

where $STD_{time}$ is the standard deviation across time, and $CONV$ is the convolution operator.

$$m_3=\max_{time}\left\{ 4.23 \log \frac{TI(D_n)}{TI(O_n)}\right\},$$

where $\max_{time}$ denotes the maximum value of the time history for each test sequence. This measure selects the video frame that has the largest added motion. Finally, the quality measure $q$ is given in terms of $m_1$, $m_2$, and $m_3$ by

$$q = 4.77 - 0.992 m_1 - 0.272 m_2 -0.356 m_3.$$

From the above equation, we can see that this metric computes the quality from a linear combination of certain objective quality measures. As we show in Section 6.6, the correlation factor with subjective data is variable from 0.73 to 0.94, depending on the distortion. In addition, it is not consistent with all the bit rates, that is it gives poor results for low bit rates.