12704_SCTE_Broadband_Nov2016_COMPLETE_lowres

technical

Figure 2 above shows an original image with a perfect SSIM score of 1. While image (d) has a better SSIM score than image (b), viewers would prefer the quality of image (b).

equal to the PSNR measurement, normalised by the amount of activity in the picture or measured area. This normalisation is what helps SSIM to be more accurate than PSNR, as higher spatial activity increases the masking of the distortion. 3 While SSIM is an improved video quality measurement compared with PSNR, it still has many limitations, an example of which is shown in Figure 2 above. Next-generation video quality assessment Ultimately, none of the existing traditional techniques for video quality assessment measure distortion accurately, limiting the capabilities of modern day compression systems. In the past, there have been trials to build a video encoder that utilises the feedback of video quality measurement to improve the encoding decisions. Those attempts have been relatively unsuccessful. Effectively assessing video quality requires going beyond distortion measurement and taking into account the HVS. While SSIM made progress on measuring video quality, it is still merely a distortion measurement that utilises spatial activity, masking what actually characterises the HVS. The advantage of SSIM is its simplicity and low computation requirements, but it falls short of achieving optimal compression. Some of the properties of the HVS that are important for creating an accurate video quality measurement model include: n Contrast Distortion: The HVS is more sensitive to changes in contrast as opposed to luminance level. When measuring distortion, this needs to be taken into consideration. The

light information entering the lens in our eyes is received by an array of rods and cones that are laid along the retina. The rods and cones measure light intensity in different visible wavelength and transmit this measurement using neurons to one or more ganglion cells. In Figure 3, a model of a ganglion cell that creates a Receptive Field (RF) is shown. The RF is fed by light receptive information that is generated by a group of rods and cones. The first type of RF, shown in Figure 3A, is called “off centre”. Its highest response is when the centre is not exposed to light and the surrounding area is exposed to light. The other type of RF is called “on centre” and is shown in Figure 3B. Its highest response is when the centre is exposed to light and the surrounding area is not exposed to light. The mechanism that responds well to transitions between lighted and non-lit areas is the main reason that our visual system is sensitive to contrast and less sensitive to pure light.

Figure 3.A: Off-centre receptive field

Figure 3.B: On-centre receptive field

Note: video compression standards are compressing the light level (luminance) information and not the contrast level. n Spatial Frequency: The HVS model uses spatial frequencies to measure visual information. The receptive fields shown in Figure 3 exist in a different physical

3 https://arxiv.org/ftp/arxiv/papers/1503/1503.06680.pdf

46

Vol. 38 No. 4 - November 2016 Issue