Video coding
 | This section may contain an excessive amount of intricate detail that may interest only a particular audience. Specifically, this is not the place to explain the general concept of video compression in such detail; focus should be kept on the H.262 video codec..(May 2020) |
Picture sampling
An HDTV camera with 8-bit sampling generates a raw video stream of 25 × 1920 × 1080 × 3 = 155,520,000 bytes per second for 25 frame-per-second video (using the 4:4:4 sampling format). This stream of data must be compressed if digital TV is to fit in the bandwidth of available TV channels and if movies are to fit on DVDs. Video compression is practical because the data in pictures is often redundant in space and time. For example, the sky can be blue across the top of a picture and that blue sky can persist for frame after frame. Also, because of the way the eye works, it is possible to delete or approximate some data from video pictures with little or no noticeable degradation in image quality.
A common (and old) trick to reduce the amount of data is to separate each complete "frame" of video into two "fields" upon broadcast/encoding: the "top field", which is the odd numbered horizontal lines, and the "bottom field", which is the even numbered lines. Upon reception/decoding, the two fields are displayed alternately with the lines of one field interleaving between the lines of the previous field; this format is called interlaced video. The typical field rate is 50 (Europe/PAL) or 59.94 (US/NTSC) fields per second, corresponding to 25 (Europe/PAL) or 29.97 (North America/NTSC) whole frames per second. If the video is not interlaced, then it is called progressive scan video and each picture is a complete frame. MPEG-2 supports both options.
Digital television requires that these pictures be digitized so that they can be processed by computer hardware. Each picture element (a pixel) is then represented by one luma number and two chroma numbers. These describe the brightness and the color of the pixel (see YCbCr). Thus, each digitized picture is initially represented by three rectangular arrays of numbers.
Another common practice to reduce the amount of data to be processed is to subsample the two chroma planes (after low-pass filtering to avoid aliasing). This works because the human visual system better resolves details of brightness than details in the hue and saturation of colors. The term 4:2:2 is used for video with the chroma subsampled by a ratio of 2:1 horizontally, and 4:2:0 is used for video with the chroma subsampled by 2:1 both vertically and horizontally. Video that has luma and chroma at the same resolution is called 4:4:4. The MPEG-2 Video document considers all three sampling types, although 4:2:0 is by far the most common for consumer video, and there are no defined "profiles" of MPEG-2 for 4:4:4 video (see below for further discussion of profiles).
While the discussion below in this section generally describes MPEG-2 video compression, there are many details that are not discussed, including details involving fields, chrominance formats, responses to scene changes, special codes that label the parts of the bitstream, and other pieces of information. Aside from features for handling fields for interlaced coding, MPEG-2 Video is very similar to MPEG-1 Video (and even quite similar to the earlier H.261 standard), so the entire description below applies equally well to MPEG-1.
I-frames, P-frames, and B-frames
MPEG-2 includes three basic types of coded frames: intra-coded frames (I-frames), predictive-coded frames (P-frames), and bidirectionally-predictive-coded frames (B-frames).
An I-frame is a separately-compressed version of a single uncompressed (raw) frame. The coding of an I-frame takes advantage of spatial redundancy and of the inability of the eye to detect certain changes in the image. Unlike P-frames and B-frames, I-frames do not depend on data in the preceding or the following frames, and so their coding is very similar to how a still photograph would be coded (roughly similar to JPEG picture coding). Briefly, the raw frame is divided into 8 pixel by 8 pixel blocks. The data in each block is transformed by the discrete cosine transform (DCT). The result is an 8×8 matrix of coefficients that have real number values. The transform converts spatial variations into frequency variations, but it does not change the information in the block; if the transform is computed with perfect precision, the original block can be recreated exactly by applying the inverse cosine transform (also with perfect precision). The conversion from 8-bit integers to real-valued transform coefficients actually expands the amount of data used at this stage of the processing, but the advantage of the transformation is that the image data can then be approximated by quantizing the coefficients. Many of the transform coefficients, usually the higher frequency components, will be zero after the quantization, which is basically a rounding operation. The penalty of this step is the loss of some subtle distinctions in brightness and color. The quantization may either be coarse or fine, as selected by the encoder. If the quantization is not too coarse and one applies the inverse transform to the matrix after it is quantized, one gets an image that looks very similar to the original image but is not quite the same. Next, the quantized coefficient matrix is itself compressed. Typically, one corner of the 8×8 array of coefficients contains only zeros after quantization is applied. By starting in the opposite corner of the matrix, then zigzagging through the matrix to combine the coefficients into a string, then substituting run-length codes for consecutive zeros in that string, and then applying Huffman coding to that result, one reduces the matrix to a smaller quantity of data. It is this entropy coded data that is broadcast or that is put on DVDs. In the receiver or the player, the whole process is reversed, enabling the receiver to reconstruct, to a close approximation, the original frame.
The processing of B-frames is similar to that of P-frames except that B-frames use the picture in a subsequent reference frame as well as the picture in a preceding reference frame. As a result, B-frames usually provide more compression than P-frames. B-frames are never reference frames in MPEG-2 Video.
Typically, every 15th frame or so is made into an I-frame. P-frames and B-frames might follow an I-frame like this, IBBPBBPBBPBB(I), to form a Group of Pictures (GOP); however, the standard is flexible about this. The encoder selects which pictures are coded as I-, P-, and B-frames.
Macroblocks
P-frames provide more compression than I-frames because they take advantage of the data in a previous I-frame or P-frame – a reference frame. To generate a P-frame, the previous reference frame is reconstructed, just as it would be in a TV receiver or DVD player. The frame being compressed is divided into 16 pixel by 16 pixel macroblocks. Then, for each of those macroblocks, the reconstructed reference frame is searched to find a 16 by 16 area that closely matches the content of the macroblock being compressed. The offset is encoded as a "motion vector". Frequently, the offset is zero, but if something in the picture is moving, the offset might be something like 23 pixels to the right and 4-and-a-half pixels up. In MPEG-1 and MPEG-2, motion vector values can either represent integer offsets or half-integer offsets. The match between the two regions will often not be perfect. To correct for this, the encoder takes the difference of all corresponding pixels of the two regions, and on that macroblock difference then computes the DCT and strings of coefficient values for the four 8×8 areas in the 16×16 macroblock as described above. This "residual" is appended to the motion vector and the result sent to the receiver or stored on the DVD for each macroblock being compressed. Sometimes no suitable match is found. Then, the macroblock is treated like an I-frame macroblock.
