DCDi by Faroudja
What is DCDi? The following excerpts from an article originally written and published by Stacey Spears, Video Editor, and Brian Florian, Editor, Canada, of Secrets of Home Theater and High Fidelity (www.hometheaterhifi.com ) in 2003 to explain the technology behind the three-time Emmy® award-winning video processing technology.
An overview of DCDi by Stacey Spears and Brian Florian
Faroudja technology is now available for the masses, embedded in consumer electronics. What was once only available in products costing $15,000 or more in large home theaters, is now in multiple products well below $1,000. How do you identify products with this technology? Simply look for the DCDi® logo. In some rare cases, Faroudja technology might be inside even when there is no logo to be found. There are several products on the market today, from DVD players to display devices, that contain this Emmy® award winning technology.
On the design side, DCDi is a specific technology Faroudja introduced a couple of years back in their broadcast up-converter. Today the DCDi logo is really used to identify a superset of Faroudja technology. This includes their patented film mode detection, bad-edit detection, cross-color suppressor, and DCDi.
What is DCDi and why do you or I care? Before we can explain that, some background information must be provided. Brian Florian will show you how film is transferred to video, which is really the heart of it all.
An explanation of film-to-video frame rate conversion for NTSC
To better understand the upcoming concepts, one must be armed with some basic knowledge of how film gets transferred to video, as well as the nature of interlaced versus progressive display. As such, the following information is not intended to be a definitive paper on the subject, but should serve as a good introduction for all.
The visuals and animations presented here provide key understanding of concepts, and will reward repeat viewing.
Motion pictures are comprised not of motion at all, but numerous stills shown in rapid succession. For the films we all watch at the theater, 24 frames are shown in one second (24 frames per second, or 24fps). The NTSC television system differs from film in this regard, making it complicated to show film on video.
Old CRT televisions created their image by drawing (scanning) lines of light on the CRT face, left to right, top to bottom, to produce a picture over the entire screen. The resultant images that make up the motion picture are comprised of two interlaced fields: that is, the first field consists of all the odd lines (1 through 525), and the second field consists of all the even lines (2 through 524). The result is that only half of the video's display is drawn every 60th of a second. A simulation of this is shown on the left. Field 1 is scanned, and then Field 2 is scanned. Traditional talk quotes NTSC television as having 30 frames per second (as opposed to film's 24), each being comprised of two interlaced fields. This is actually misleading: The NTSC interlaced system shows 60 unique images per second, but each one uses only half of the vertical resolution available on the display. Only if the source material contained 30 unique frames per second could you say that two fields form a single frame but in reality, video material such as the evening news is true 60 fields per second. So we don't want to think of interlaced televisions in terms of frames but rather in terms of fields, interlaced fields, and 60 of them per second.
The principal drawbacks of an interlaced display are (A) visible line structure, (B) flicker caused by the rapid alternating of the fields, and most important, (C) artifacts such as 'feathering' (also referred to as 'combing') and 'line twitter'. Visual artifacts like these last two occur anytime the subject or the camera is in a different position from field to field. The subject will be in one position for one field, and in another position for the next, resulting in jagged edges (feathering) or shimmering horizontal lines (twitter).
The animation on the left shows an example of an interlaced display trying to show a tomato moving from left to right. Each field shows the tomato a little farther to the right than the previous. Because the fields are interlaced, jagged vertical edges can't help but exist, except during for the last two fields (5 and 6) where the tomato is stationary. The further back you are from an interlaced display (or the smaller the display is), the less this and other artifacts are noticed. If you want to see the effect in real life, just stick your nose up to an interlaced TV. Focus in on an objects edge that is stationary and wait for it to move. You will notice the problem right away.
At left is an interlaced image of a skier. Not only is the flicker annoying, but have a good look at the ski-pole: It comes and goes because its so fine it can only be found in one of the two interlaced fields. This is line twitter. This artifact manifests it self when fine detail is less than 2 scan lines high. It is exasperated during vertical movement as the fields alternate. Often fine detail is filtered before being encoded to minimize these artifacts when played back at home on your interlaced display device. Because of this, we have yet to experience the full potential of DVD.
Transforming film to video
The preceding basic knowledge of interlacing is necessary to understand the transfer of film to video, because it is an important factor in what we end up seeing.
Motion picture photography is based on 24 frames per second. Time to call to mind all that math you learned in school and realize that 24 doesn't go into 60 very easily. To boil it down a little, our challenge is to make 4 frames from the film fit as evenly as possible across 10 video fields. We can't just double up the fields on every fourth film frame or we'd get a real 'stuttered' look. Instead, a process is used known as 3-2 pulldown to create 10 video fields from 4 film frames. This form of telecine alternates between creating 3 fields from a film frame and 2 fields from a film frame. Hence the name 3-2.
Consider now our flow chart of the 3-2 pulldown performed on four frames of this movie scene:
Pretty cool right? It is and it isn't. 3-2 pulldown inherits much of the artifacts we described when talking about interlaced video. Anytime a field follows one made from a different film frame (noted above by the "!" icon), there exist the possibility for anomalies in what we see, feathering and twittering being great examples. Absolutely any differences between the two film frames that make up the video frame (the last field of one frame and the first field of the next frame), be it brightness, color, or especially motion, are going to result in some artifact as the two fields merge on screen. Even our little animated synthesis of the final interlaced product, which actually contains 10 interlaced pieces, shows evidence of such anomalies as the flying police cars move ahead. Such is life.
HDTV progressive displays
Progressive displays, such as high-definition plasma and LCD TVs and DLP/D-iLA projectors, can only show progressive scanned images as opposed to interlaced. In order to do this, the display must scan at a higher rate, 2x the speed of NTSC. Because we are scanning at twice the speed, we can draw an entire frame in the same amount of time it takes an interlaced system to draw a single field. We learned above that an interlaced display shows 60 fields per second. But with progressive, each "field" is now a complete picture including all scan lines, top to bottom, so we will now call it a frame, and we are showing 60 of those per second (written as 60 Hz). [Of course, only 24 of those are unique if the source is film based.]
The benefits of a progressive display are no flicker, scan lines are much less visible (permitting closer seating to the display), and they have none of the artifacts we described for the interlaced display, as long as the source material is progressive in nature (film or a progressive video camera).
Movies on DVD are almost always decoded as interlaced fields yet all of the film's original frames are there, just broken up. What we're going to talk about next is how we take the interlaced content of DVD and recreate the full film frames so we can display them progressively. The term commonly used to restore the progressive image is deinterlacing, though we think it is more correct to call it re-interleaving, which is a subset of deinterlacing.
Deinterlacing (or re-interleaving) involves assembling pairs of interlaced fields into one progressive frame (1/60 of a second long), and showing it at least twice to use up the same amount of time as two fields. The need for 60 flashes on the screen each second stems from a biological property called the Flicker Fusion Frequency, meaning how many flashes that we need to see each second so that we (our brains) fuse the image into one where we don't see a flicker.
For every film frame that had three fields made from it, the third field is a duplicate of the first, and (if the MPEG-2 encoder is behaving properly) won't even be stored on the DVD. Instead of encoding the duplicate fields, the DVD flags repeat_first_field and top_field_first are used to instruct the MPEG decoder where to place these duplicate fields during playback.
The progressive output of a DVD player should assemble 2 fields from each film frame and create a complete progressive one that looks just like the original film frame. You should now be thinking that the DVD will once again have 24 frames to show in one second. But the progressive display is still expecting 60 complete frames per second. In order to space them out, the DVD player shows the complete frames in this order: 1, 1, 1, 2, 2, 3, 3, 3, 4, 4 and so on.
This form of display gives us a moving image very close to the original film. It has a tendency to "judder" a bit though, as every other film frame lasts 1/60 of a second longer than the previous one. Even our little synthesis of the final product, which actually contains 10 pieces, shows this judder. In the future, both the player and the display could increase their display rate above 60 fields per second, to 72 per second. At that point, the fields would only last 1/72 of a second, permitting the player to show every film frame three times (24 x 3 = 72), eliminating the motion judder, and also helping us with the Flicker Fusion Frequency problem (60 flashes per second are just barely enough in a well lit viewing environment). This would look like: 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4 and so on. 72 fps will only work with film based sources though, as it is a multiple of 24. It will not work well with video sources which are 60 field per second.
The re-interleaving process we've just covered is specific to 24fps film material which is MPEG-2 decoded (as interlaced fields). It's really a matter of putting the right fields together so it's fairly simple. Deinterlacing native NTSC interlaced video material is much more complicated. In such video material, each field is a unique image in time, and in order to be deinterlaced at an acceptable level, it requires getting into motion-adaptive and motion-compensation algorithms to overcome the inherent problems of the interlaced material. There is no best method, and the two mentioned are expensive to implement.
(Note: NTSC does not really run at 60 Hz; it is technically 59.94 Hz. The industry rounds it up to make it easier to read. If you did play back video at 60 Hz instead of 59.94 Hz, you would end up with a dropped frame approximately once every 20 seconds.)
Now that you have a basic understanding of how film is transferred to video, a few more terms need to be explained before we move on. We like to use the term deinterlacing, while others may call it line doubling or even I to P conversion. (Where I means interlaced and P means progressive). All of these really mean the same thing. We will also talk about vertical resolution. As Brian explained above, there are 525 horizontal scan lines (see diagram below) that make up an NTSC image. These horizontal lines are how we measure the vertical resolution.
We will be using the terms video mode and film mode to describe the type of deinterlacing algorithm used. When we say Film mode, we are to the algorithm that will detect the 3-2 pulldown cadence and weave the two fields of video into one that would match the original frame of film. When this is done, you are using all 525 original lines, which gives you the full vertical resolution of the image. This means even scenes that contain fast motion will be displayed in full resolution.
When we say Video mode, we are referring to the algorithm that uses interpolation to create a full 525 line image. The most basic form of video mode is where one field is used. If the field that we currently have is the one that contains the odd number of lines like 1, 3, 5, etc., then interpolation will be used to create the even lines like 2, 4, 6, etc. To create line 2, we would perform some type of average of lines 1 and 3. When only interpolation is used, you no longer have the full vertical resolution. In scenes where fast motion exists, you lose half of the vertical resolution. A more advanced form of video mode is called motion-adaptive, which is what Faroudja employs. This algorithm will not only interpolate, but will also try and weave together two fields, which would provide more vertical resolution. On a pixel-by-pixel basis, this algorithm will weave together areas that are static (no motion) and interpolate when there is motion.
We will break down some of the Faroudja processing into its various parts. These include film mode detection/bad edit detection, and DCDi.
Film mode / bad edit detection
In 1989 Faroudja invented and patented film mode detection (also called 3-2 pulldown detection). At that time, Faroudja was the only company who had the ability to detect the original frames of film within the video stream and reconstruct an accurate image. This yielded an image that was free of motion artifacts, and it contained the full vertical resolution of the image.
In the early 90's, a video magazine conducted a couple of high profile video shootouts. They obtained all of the current line doublers, as they were called back then, and put them head-to-head. One product stood out from all the rest, the Faroudja LD-100. This was the only product that had film mode detection, and rightly so, as Faroudja held the patent on it. This processor was not just a little better than the other, it was a lot better.
A couple of years ago, other companies began to introduce video products that also had film mode detection. This narrowed the performance gap between Faroudja and the rest. However, like any great company, Faroudja did not sit around reminiscing about the good old days, they continued to invest in research and development, and DCDi was born. In fact, DCDi just won Faroudja another Emmy® .
Just having film mode detection is no longer just good enough. Why? Because the 3-2 pulldown cadence, like the world, is not perfect. Problems occur which cause that cadence to break. You might get a 2-2 or 3-3, or 4-1 cadence to name just a few of the possibilities, and this can confuse some of the other technology in use today. Enter bad edit detection. Errors can and do happen at any stage between the time the images travel from the film to your TV screen. These errors will show up as artifacts on screen. The most common artifact, a comb, happens when the video processor combines two fields of video that come from two different frames of film. Bad edit detection is able to see this problem coming and avoid it all together. Figure 1 below shows an example of what a comb would look like on-screen. Notice the picture looks like someone ran a comb through it, thus the name.
How is the problem avoided? By switching between film and video deinterlacing. If you can't combine two fields together because they don't belong, the product has to interpolate. Hopefully this only happens for a few fields of video. All deinterlacing algorithms switch between film and video, but the strength lies in how quickly you can detect the error and switch. Many deinterlacers switch after it is too late. The goal is to switch to video mode before you display an artifact and switch back to film mode as quickly as possible. Remember, video mode means it is dealing with a signal that originated as video, with 60 fields (equivalent to 30 frames) per second. Film mode means the original was a film, with 24 frames per second. The problem occurs because the final viewing medium is a TV, with 60 fields per second, regardless of what the original source was.
DCDi is a video mode algorithm that stands for Directional Correlation Deinterlacing. It was designed for video based material like fast-paced sporting events. Its purpose is to eliminate jagged edges (jaggies) along diagonal lines caused by interpolation. If you remember, you are not simply weaving together two fields of video that match, you have to create new information through the art of interpolation which is really a fancy way to say you are guessing. DCDi monitors edge transitions and fills in the gaps. The technology was introduced a few years back in the digital format translator, a $50,000 system that broadcasters like CBS use to upconvert NTSC to HD. It was/is used to upconvert standard definition material (480i, what we have on conventional TV) to enhanced definition quality (480p). It is also used to deinterlace 1080i to 1080p HD material. You may already being enjoying DCDi today on your digital TV.
In the figure above, you will see how DCDi makes the Stars and Stripes much more dramatic, and it is really a terrific illustration of how powerful DCDi is. On the left is the original image on a TV. The flag is blowing in the wind, and this is a very tough image for a TV to show. On the right are enlargements of an area in the middle of the picture. At the top, right, is that enlarged area of the flag, with DCDi turned off. Notice the junctions of the red and white stripes. You can see jagged lines. With DCDi turned on (bottom, right), the jagged lines are gone, and the junctions between the red and white stripes are smooth. This is a huge technical accomplishment by Faroudja engineers.
Because DCDi is a video algorithm (an algorithm is a series of mathematical formulas), you might wonder how it affects viewing a film on TV. Remember, in order to avoid artifacts, a video processor will switch modes (film vs. video - video vs. film). If the transition between video and film is not done properly by the studio, it is called a Bad Edit. The video processor will then treat the film material as video during those sections of bad edits. There are a couple of giveaways when the processor has switched from film mode to video mode. First is the loss of resolution. This is minimized because the Faroudja algorithm is motion adaptive. The second is the appearance of jaggies along diagonal edges. DCDi hides a good portion of the jaggies, so you never realize when it changes from and to film mode, which is the whole point! DCDi makes the movie watching experience more enjoyable because the annoying artifacts are all gone, so all you have to worry about is whether there is any more microwave popcorn in the kitchen cabinet.
Implementing the technology
The Faroudja DCDi technology is found in most of the TV SoCs offered by ST. Click here to link to the STMicroelectronics product pages.
"The Fifth Element" Copyright 1997, Columbia Pictures
"Casablanca" Copyright 1942, Warner Brothers
"Top Gun" Copyright, 1986, Paramount Pictures
"Galaxy Quest" Copyright 1999, Dreamworks