Post a Comment
Feature
Video quality: A hands-on view
Deinterlacers all may look the same on paper, but engineering intuition suggests that they probably do not produce identical results with real-life video material. EDN takes a look at the contenders and the pretenders.
By Brian Dipert, Technical Editor -- EDN, 6/7/2001
Video deinterlacing and scaling, like many other aspects of multimedia, are deceptively simple in theory but challenging to properly implement when you face imperfect and unpredictable real-life images (Reference 1). Full-featured deinterlacers detect and reconstruct computer graphics and NTSC- and PAL-encoded film, progressively scan convert-video fields, handle a mixture of film and video content within the same frame, and tailor their output resolution and aspect ratio to the chosen display device. The numerous techniques for handling these various conversions each have a corresponding set of cost, quality, and other trade-offs.
Predictably, in a situation such as this, the proponents of each method emphasize its strengths and minimize or even 'forget' to mention its shortcomings. Every time I encounter such a state of affairs, I'm tempted to do some hands-on analysis of my own. Table 1 outlines the equipment I've used in this video study. And the sidebar 'Endless views' directs you to the additional research results.
Visual variables
Before diving into the details, here are a few disclaimers. Before I began in-depth research of video technology, I couldn't tell the difference between the composite-video and S-Video outputs of the DVD player connected to my 31-in. interlaced television. Once I knew what composite-filtering artifacts to look for, they became glaringly apparent. And, after having experienced a progressive-scan, component-video-driven, high-quality display during this project, I'm already scheming (or, more realistically, dreaming) about how I can convince my spouse that we need to upgrade our DVD player and TV, as well as the video-switching home-theater receiver that links them.
Getting a taste of deinterlaced-video quality also motivated me to put an antenna on the roof, borrow a DTV receiver and a few examples of PC-based DTV hardware and software, and tune in to the digital television content that's slowly appearing in the Sacramento, CA, broadcast area. My experiences aren't unique, and they hold clues as to the realistic rate at which consumers will migrate to higher quality video. Until now, I'd been perfectly content with NTSC television, because I hadn't experienced anything better. Yes, I'd seen and appreciated the HDTV and line-doubled NTSC demos at the Consumer Electronics Show and other conferences, but my memory of them had faded by the time I got back home.
The benefits of a progressive-scan display didn't really stick until I compared my television and a progressive-scan monitor side by side. Whether you're evaluating speakers, displays, or lossy multimedia compression algorithms, your ears, eyes, and brain often compensate for imperfections, leaving any one product sounding and looking fine when not in the presence of a clearly better alternative (Reference 2). Even with this high-quality monitor, I can't always see all of the deinterlacing characteristics that the vendors' literature touts. Improper field-combination artifacts aren't noticeable, for example, except on carefully crafted test patterns.
Perceived video quality is a function of numerous variables, including the inherent quality of the source. Are you watching a 4:4:4 component video stream or high-resolution computer RGB output, or are you watching a noisy, RF-modulated NTSC composite video broadcast? What type of display are you using to observe the video source? Is it a cheap, 13-in. interlaced television, or a $100,000 front-projection system with a huge screen? Ironically, the more money you invest in your video equipment in search of absolute image perfection, the more noticeable the artifact impediments to that perfection become. A 'soft' resolution image on a small screen does an effective job of hiding 'feathering,' aliasing, analog and digital noise, DCT-pixel-matrix blocking, and other defects.
What are the ambient lighting conditions under which you're observing the images, and how close are you to the display? How sharp is your vision, and how much experience do you have searching for image artifacts (Reference 3)? Feedback I received from some readers on earlier digital-audio articles indicates that they didn't understand my intent of that meticulous study: to discover how the codecs worked, not to make quality conclusions about them. When I
Great gobs of gear
I used Toshiba's SD-6200 (Figure 1a) as my video source, and it's also one of the deinterlacing alternatives I evaluated. This DVD player outputs interlaced composite video, S-Video, and component video. The two-chip Genesis Microchip set inside it (gmVLX1A-X video processor and scaler, and gmAFMC film mode processor) also enables it to output progressive video over the component connections.
On the other end of the video chain is Princeton Graphics' AF3.0HD (Figure 1b), a 32-in. (30-in.-viewable), wide-screen HDTV monitor. The AF3.0HD contains a simple 'bob'-mode line doubler that operates when you employ the monitor's composite-video and S-Video inputs. The RGBHV (red-green-blue-horizontal vertical sync), VGA, and two sets of component video inputs bypass the monitor's doubler, luma/chroma separator, color decoder, and other intermediary video- processing circuits (Reference 5). Silicon Graphics has provided its model 1600SW 1600´1024-pixel, wide-screen-LCD flat-panel monitor and MultiLink adapter (Figure 1c), which I plan to add to my imaging testbed. The LCD will enable me to more precisely observe output resolutions and various video artifacts, but I'll need to carefully ensure that the 1600SW's internal scaler and the MultiLink's interface and aspect ratio converters aren't the root cause of any observed image degradation.
Several deinterlacers in the lab sit between the DVD player's interlaced outputs and the Princeton Graphics monitor's component and RGBHV inputs. Silicon Image bases its iScan Pro (Figure 2a) on the SiI503 chip. The iScan Pro includes composite, S-Video, and component inputs; accepts NTSC, PAL, and SECAM video formats; and outputs progressive-scan RGB and component video. Focus Enhancements' QuadScan Elite (Figure 2b) contains the same two-chip Genesis Microchip set as does the Toshiba DVD player, and it can scale interlaced component, S-Video, and composite inputs all the way up to a 1365´1024-pixel output resolution.
Three vendors provided evaluation boards. The NV221 demonstration board from nDSP mates over VME to a front-end video-ADC board containing a Micronas VPC3230 video processor. Sage's EVL9200TV board showcases the company's FLI2000S video decoder, FLI2200 digital-video deinterlacer (Figure 2c), and FLI2220 digital-video enhancer, along with triple 10-bit video DACs. Both the nDSP and Sage boards can operate in either stand-alone mode or with PC-parallel-port control. So far, I've run both boards in stand-alone mode and have manipulated their functions via DIP switches.
Genesis Microchip also provided me a reference board for their brand-new gm6010 video processor. Its onboard tuner would have enabled me to test its performance not only with composite, S-Video and component inputs, but also with noisy terrestrial and cable-broadcast video signals. However, its subpar deinterlacing results on DVD-player-generated signals indicated immature firmware development. Because competitors' devices are older and therefore offer more fully debugged algorithms, I decided to defer publishing gm6010 testing data. Keep an eye on the
To ensure that each deinterlacing alternative had the chance to present itself in the best possible light, I employed high-quality video cables for all connections, including an UltraLink cable with an HD-15 connector on one end and five BNCs on the other end. I viewed the CRT from approximately two feet away. (Hey, it's a small office!) Keep this unnatural viewing position in mind as you analyze the information that follows, and please don't tell my mother that I'm
Obvious observations
I first hooked up the DVD player's composite, S-Video, and component outputs directly to the television. I happened to have a copy of the movie
From my theoretical research, I expected to see a less significant quality improvement when switching from S-Video to component video versus the already-observed composite-to-S-Video transition. I was surprised, therefore, when I switched the AF3.0HD over to a mode in which the DVD player's progressive-component outputs were driving it. The images on the screen seemed to jump out at me with enhanced color richness and sharpness that even revealed flaws on the film used to create the DVD master. What happened?
For one thing, I was viewing source material that was stored in 4:2:0 component-video format. Instead of running that component video through multiple bandwidth-reducing format transformations to S-Video or composite video and then transforming it
The DVD player's setup menu also had edge sharpening enabled in the default settings for the component-video outputs. Sharpening can improve perceived image quality, particularly at extended viewing distances, but when you use it to an extreme, it can inject noise and other undesirable artifacts into the picture. Like many other aspects of video, sharpness is a matter of taste, so for my testing I either set it to a nominal value if I had an 'analog' range of options available, or I disabled it if the 'binary' options were more limited.
Film...video...action!
For my next set of tests, I turned to the 'Montage of Images' section of the
After repeatedly playing 'Montage of Images' through various pieces of deinterlacing equipment, using all possible interconnections and from my intimate vantage point near the screen, I was surprised by how
I don't fault Princeton for the company's line-doubling choice. The set, which has been in production for several years, costs thousands of dollars, and at the time it was in development, robust deinterlacing technology would have further increased the price by hundreds or thousands of dollars. Today, thanks to semiconductor integration, that same level of deinterlacing technology costs only tens of dollars or less. Exemplifying this trend, Princeton Graphics' new AI3.6HD 4:3-aspect-ratio monitor includes an SiI503 deinterlacing processor.
To use an analogy, many PC manufacturers integrate a basic level of 2- and 3-D-graphics function through the core logic and provide options for more robust graphics capability via AGP slot expansion. Similarly, consumers can supplement or even override the TV's basic line doubling by deinterlacing at each video source, including the DVD player, or at a common midpoint, as Focus and Silicon Image do with their video-processing boxes and nDSP and Sage do with their evaluation boards. Exemplifying the option of deinterlacing at each video source, the Toshiba DVD player delivered the cleanest, sharpest picture. The player has access to and passes along to the TV the highest-quality component video. The fields it deinterlaces have not yet seen a digital-to-analog conversion, and the player's deinterlacing processor can reference the DVD's control flags to further guide its efforts.
'Sharp' doesn't necessarily mean 'best,' though. The Focus Enhancements, nDSP, Sage, and Silicon Image deinterlacers all seemed to deliver a slightly softer picture than the Toshiba-generated picture. (Further analysis will attempt to quantify this difference via lines-of-resolution measurements.) However, when their algorithms partially or fully failed a deinterlacing test, the 'feathering' or other artifacts that resulted were less objectionable than those from the Toshiba. And the aliasing artifacts resulting from diagonal edges in motion were also less objectionable with the softer presentation to obscure them.
Why were the various deinterlacers so similar in their results with film- and video-based material? Part of the reason was undoubtedly the material used to evaluate them. 'Video Essentials' is a well-known video-testing and -setup disc, published in DVD format several years ago. It has existed in laserdisc format even longer. It wouldn't surprise me if 'Montage of Images' was one of the tests the manufacturers insisted that their units 'pass' before full production. Badly edited material and movies with incorrect flag encoding, such as
Part of the reason for the similarity might also have been the testing method I chose. I approached the equipment as a consumer would, playing the DVD at normal speed. In other evaluations I've seen, reviewers single-step through the DVD and count the number of fields it takes for the equipment to change to the proper deinterlacing mode when switching from film to video or from video to film. I don't deny that there's a small but lucrative videophile segment of the buying public that will select equipment, regardless of its price, based on a one- or two-frame-delay difference. I'd postulate, though, that most consumers would barely—if at all—perceive such a delay in the context of a 60-frame/sec presentation, and price therefore becomes the primary differentiator.
My close proximity to the screen made for a worst-case viewing position, and I'm confident in my vision, at least when I'm wearing glasses. I'm also confident that the AF3.0HD was able to reproduce the entire 480-scan line vertical resolution. However, some deinterlacing artifacts are subtle and become increasingly noticeable with additional viewing; my job unfortunately doesn't allow me to spend hours watching movies! Also, the switch from one video-equipment arrangement to another took a few seconds' delay at best as I toggled between inputs using the TV's front-panel controls. More often, the switch took a few minutes and involved swapping hardware, software, and cabling. Memory and, with it, discernment, fades during these extended delays.
Torture tests
In the hope of observing more significant deinterlacing differences in situations that 'Montage of Images' might not have encompassed or that my untrained eyes might not have noticed, I next turned to a series of four video clips, three of which came from MadOnion.com's Video2000 benchmarking software. The
Deinterlacers that use field-based motion adaptivity should sense the logo motion and switch to intrafield interpolation with the same flickering display result. Deinterlacers with per-pixel motion adaptivity, however, correctly merge fields—that is, employ temporal interpolation, or 'weave,' in the stationary areas while using spatial interpolation in the moving areas. The ideal result is a stable, flicker-free image with all lines visible on the screen. The ideal rotating logo also looks clean without any feathering artifacts at its edges.
The
Deinterlacers that always weave the odd and even fields produce feathering along the edges of the pendulum. Motion-adaptive deinterlacers employ spatial interpolation around the pendulum. If they use a constant-proportion interpolation algorithm, however, you'll see visible aliasing along the pendulum edge, especially at shallow angles.
The
Ideally, the deinterlacing processor should have acquired the pattern by the time the square box changes from outline to solid white at approximately 50 fields into the sequence. Losing the telecine pattern causes the horizontal lines to flicker. The image contains little motion, because most of it is a stationary black background. The rotating logo, however, produces complex motion, making it difficult for a film-mode detector to find the 3:2 pattern.
The final 'torture test' demonstrates content that mixes video and film. This sequence consists of an unrolling roll of toilet paper, shot on film at 24 frames/sec and converted to video using 3:2 pulldown, along with superimposed text that is interlaced video and therefore contains no telecine information. Because most of the image is film, the image can cause deinterlacers that support film mode to employ an inverse-telecine algorithm for the entire frame, a choice that causes severe feathering around the text.
The proper approach to this content is to use a video-deinterlacing algorithm on the frame, which may not produce the optimal progressive-scan conversion of the film regions but results in readable text. Appropriate handling of content that mixes video- and film-based moving objects within a frame is important because entertainment material, such as music videos and commercials, increasingly employs this technique.
To create the images in Figure 3 and Figure 4, I captured still frames using PC-based DVD software. Software limitations cause the artifacts you see in the images, such as the pendulum double-image in Figure 4b, and they are not present in the properly deinterlaced display.
Results and resolution reminiscing
Table 1 shows the results for each deinterlacer, for which I tested every possible combination of analog video-input modes, including composite, S-Video, and—if supported—interlaced component, and video-output modes, including component and RGBHV. The deinterlacing results for all mode combinations were identical; therefore, I didn't further subdivide Table 1 by input and output format. Had I also tested the equipment with a randomly noisy video feed coming from my cable connection or from a terrestrial-broadcast-antenna source, I suspect I'd have seen more variability both between units and between tests on the same unit.
A red box in Table 1 indicates that the unit consistently created objectionable artifacts, and a green box indicates that the unit consistently 'passed' the test. A yellow box indicates either that the unit occasionally failed the test at certain points within the frame or that the artifacts it produced were judged less objectionable than others I'd seen from units that used a different deinterlacing method.
In analyzing the results, look first at the situations in which I forced the unit into a specific deinterlacing mode: film, video, or still video if available. These tests indicate whether the hardware and corresponding software algorithms inside the unit can properly deinterlace the material. It's obvious that the Focus Enhancements QuadScan Elite and the Toshiba SD-6200 share the same Genesis Microchip deinterlacing subsystem. The Toshiba DVD player's 'video' mode appears to use a blend of the Focus Enhancements deinterlacer's 'video' and 'still-video' modes.
Next, look at the 'auto' deinterlacing modes. Poor results here might indicate a software issue, which a firmware upgrade could fix. Such an issue could be, for example, that an algorithm 'guessed' wrong when presented with a complex blend of still and in-motion objects or of film- and video-sourced material. Alternatively, the errors might reflect a more fundamental hardware limitation; for example, the chip might have limited buffer memory that forces it to select 'film' or 'video' mode based only on analysis results of the first few scan lines in each field. The fourth torture test is ideal for assessing this situation because the video-sourced caption location is in the bottom of the frame.
Unfortunately, I had access only to the Silicon Image iScan Pro's 'auto' mode, so I wasn't able to fully assess the deinterlacing capabilities of the SiI503 chip inside. Also, the mixed film-and-video green-box result for the Princeton Graphics AF3.0HD is deceptive; because the unit's line doubler offers no film mode, there was no chance that it could fail the test.
One piece of data that Table 1
However, pixel-specific selection of temporal or spatial interpolation and of specific percentages of nearby pixels' information for this interpolation comes with an important caveat: The more pixel-specific the selection criteria, the higher the potential that random analog or digital video noise or other extraneous artifacts will fool the algorithm into 'thinking' that, for example, motion exists when it really doesn't. I noticed no incorrect decisions with the FLI2200 and the pristine DVD material I fed it, but I'd like to repeat my testing with less forgiving video content.
I have been critical of video subsystems that enable a user to stretch and horizontally distort 4:3-aspect-ratio video material to fill an entire wide-screen display (Reference 1). In fact, some televisions automatically perform this function when you drive their RGBHV or progressive-component inputs. Some of these sets don't even allow you to disable stretching, necessitating circumventing modes in deinterlacers, such as the 'squeeze' function that the Silicon Image iScan Pro offers.
However, the distortion that horizontal stretching causes was less obvious than I'd feared, particularly after my eyes and brain had a chance to compensate after a few minutes' viewing. Both the AF3.0HD and 1600SW support automatic stretching, and I didn't even realize that the AF3.0HD was distorting the 4:3-aspect-ratio torture-test clips until I viewed them on my NTSC television. Similarly, after using my desktop computer hooked to the 1600SW for a while, text looks abnormally tall when I view it on my notebook computer's 4:3 LCD.
| I've barely scratched the surface of covering the approaches and trade-offs of video technology. The following lists the topics that I hope to explore in coming months. Look for the results both in a Web-site addendum and in future print articles. I hope to:
The HDT-2000F isn't the only DTV tuner available. TeraLogic's Janus evaluation board, along with AccessDTV, and Hauppauge boards based on TeraLogic's TL880, perform tuning, transport-stream demultiplexing, and video decoding all in hardware. The Broadlogic DTA-100 card relies on demultiplexing and decoding software from Intervideo, National Semiconductor, and Ravisent Technologies. Speaking of the PC, I'd like to:
I welcome your feedback on how I prioritize these tests, as well as which tests I may have overlooked. Contact me at bdipert@pacbell.net. |
| For more information... | ||
| When you contact any of the following manufacturers directly, please let them know you read about their products in EDN. | ||
| Focus Enhancements 1-978-988-5888 www.focusinfo.com Enter No. 356 | Genesis Microchip 1-905-889-5400 www.genesis-microchip.com | nDSP 1-408-626-1640 www.dsp.com |
| Princeton Graphics 1-714-751-8405 www.princetongraphics.com | Sage 1-408-383-5300 www.sageinc.com | Silicon Graphics 1-650-960-1980 www.sgi.com |
| Silicon Image 1-408-616-4000 www.siimage.com | Toshiba 1-212-596-0600 www.toshiba.com | Other companies mentioned in this article
|
Author Information
Contact Technical Editor Brian Dipert at 1-916-454-5242, fax 1-530-937-8147, e-mail bdipert@pacbell.net.
REFERENCE
- Dipert, Brian, Video improvements obviate big bit streams,
EDN , March 15, 2001, pg 83. - Dipert, Brian, Security scheme doesn't hold water(marking),
EDN , Dec 21, 2000, pg 35. - Dipert, Brian, Digital audio gets an audition. Part two: lossy compression,
EDN , Jan 18, 2001, pg 87. - Munsil, Don, and Stacey Spears, 'DVD Benchmark: Progressive DVD Player Shootout - December 2000,'
Secrets of Home Theater and High Fidelity , www.hometheaterhifi.com/volume_7_4/dvd-benchmark-progressive-shootout-12-2000.html. - Dipert, Brian, A crash course in color conversions,
EDN , June 7, 2001, pg 46. - Dipert, Brian, Digital photography clicks,
EDN , Dec 18, 1997, pg 50.
ACKNOWLEDGMENTS
I'd like to thank all the companies that have provided hardware, software, video material, and consultation for this year's video research: AccessDTV, BroadLogic, Creative Labs, CyberLink, DisplayMate Technologies, EnTech Taiwan, Focus Enhancements, Genesis Microchip, Hauppauge Computer Works, Intel, InterVideo, Joe Kane Productions, Kayye Consulting, MadOnion.com, Matrox, MGI Software, Microsoft, National Semiconductor, nDSP, Nvidia, Ovation Software, Panasonic, Princeton Graphics, Ravisent Technologies, Sage, Sencore, Silicon Graphics, Silicon Image, TeraLogic, Terk, 3Dlabs, Toshiba, Ultralink, and Unapix Entertainment.
