Digital Video - November 2007 - (Page 50) MASTER&COMPRESSOR >KEY COMPRESSION LINGO BANDING This is an artifact typically caused by either poor quality 10-bit-to-8bit conversion (preencode) or coarse quantization in the encode process. MACROBLOCK In most video compression systems, a frame is divided into groups of 16x16 pixels. Each of these is subdivided into smaller squares, from 8x8 to as little as 4x4. A full 16x16 macroblock is the basis of motion estimation in MPEG-2 inter-frame encoding. PSNR Peak signal-to-noise ratio, expressed in decibels. In the Sonic Cinevision HD disc encoder, luminance intensities throughout an image are compared between the original source and the compressed result. With low PSNR, details get lost. A PSNR of 40db or higher is considered to be very good. This is a useful metric, but studies show that the best judge of image quality is a human viewer watching a video play back. QUANTIZATION There are entire chapters in compression textbooks devoted to various quantization methods, so I will define it in the most basic terms. Quantization is the lossy part of a compression system. If it works properly, it uses the smallest quantity of numeric values needed to describe the range of characteristics of an image, without being noticeable to the average viewer. In other words, it is a system for approximating a value, rather than attempting to precisely describe it. Imagine if your TV meteorologist gave out the daily highs and lows in your city rounded to within three degree-increments. For example, instead of saying that today’s high will be 90, they’ll say 87. You probably wouldn’t notice the error when you go outdoors after watching the weather report—particularly on a very cold or very hot day. However, if you are cooking a turkey in an oven, you need a precise scale of temperature in order to ensure it is safe to eat. In the case of the weather report, the quantization level is coarse, or high. In the case of cooking a turkey, you need much finer quantization. In video encoding, a high/coarse quantization uses up less bits than a low/fine quantization, and therefore may have artifacts if it has been incorrectly applied. For example, encoding several frames of black should not require fine quantization, but encoding a color gradient could. In any case, you can save a lot of bits when quantization is applied correctly. —D.O.W. Left: This is the Macroblock class control window. The analysis in the bottom area shows that 20.23 percent of all the macroblocks in the image are textured areas, with a high (bad) QP of 6. To determine where to add bits, you would look at Luma only next. Right: A look at the Y channel, or luminance only. Eighty-four percent of the image is textured with a quality of 6. One fix would be to throw more bits into the Luma channel, where the human eye sees the most detail. reencoding. CineVision 2.0 offers fewer problematic segments. reencode parameters for VC-1, but it is more Once you’ve done this, the segments are flexible than PSE, handling the three different now considered to be “takes” by codecs with very high-quality results. CineVision. You can encode segments mulHowever, if you do a lot of VC-1 work—and tiple times, using different parameters, until it has to get done really fast, with the most you get the best looking one. CineVision precise control—PSE is the way to go. allows you to compare takes to ease the To further analyze quantization, PSE decision process. looks at macroblocks in each of the three Now you need to adjust the settings for color channels used in the component YUV this segment. To improve the look of this color space in VC-1: Y (luminance), U and V shot, you might need to increase the adap(chrominance). Within each tive quantization strength, Discover further HD of these color color channels, and select a mode such as codec info. PSE further subdivides into brightness or detail, to Visit DV online. four characteristics: smooth improve the quality of the areas, textured areas, dark/smooth, and image. Once you’ve done this you can mixed. Having this level of detail helps us examine the results in the Timeline window. determine where the good and bad quantiNow let’s look at VC-1 encoding with zation is happening in a selected segment. Sonic’s new Parallel Stream Encoder (PSE), designed for enterprise-class VC-1 encoding with help from Microsoft. Compared to CONCLUSION CineVision 2.0, PSE takes a different approach Quality HD encoding isn’t always easy. The to compression in both hardware and softcodecs themselves are still in their infancy, ware. In terms of hardware, PSE is optimized and yet, the encoding tools are much more for multiple “blade” servers, each of which sophisticated than in standard DVD. can be assigned a section or chapter of a clip. Hopefully in the future less human intervenEach individual blade can run on a dual or tion will be needed for high quality comquad-core CPU, where each core is capable of pression. For now, when faced with massive working with one section of the frame. For amounts of content and restricted disc example, with a dual-core, each core can be space, you’ll need to take the time to learn configured to handle half of the frame. tools like Sonic’s CineVision. If you really Contrast this with the CineVision, where only want to understand how a codec works, this a single computer, with eight cores or more, is is the way to go. DV best suited for encoding any of the three compliant codecs for Blu-ray or HD DVD. David O. Weissman is general manager of From the software side, PSE exposes a Video Arts (www.vidarts.com), an interacvery wide array of parameters for segment tive digital design studio in San Francisco. 50 dv november 2007 www.dv.com http://www.dv.com http://www.vidarts.com http://www.dv.com
For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.