Video Compression Bee Fong

Lossy Compression  Inter Frame Compression Compression among frames Compression among frames  Intra Frame Compression Compression in a frame Compression in a frame

Intra Frame Compression  Similar to Image Compression Do not discussion, please refer to the image compression section (JPEG) Do not discussion, please refer to the image compression section (JPEG)

Coding Order != Display Order  Figure 1  Figure 2 Encoding rate may be slow, but Decoding rate is fast

Inter Frame Compression  I Frame Intra Frame Intra Frame  P Frame Predictive Frame Predictive Frame  B Frame Bidirectional Frame Bidirectional Frame  D Frame

Compression Methods  Subsampling Duplicate frames, 1 compression stream Duplicate frames, 1 compression stream  Differencing Compare a Frame with its predecessor. Take the difference. Compare a Frame with its predecessor. Take the difference.  Block Differencing Frame divided into blocks, compare block difference. Frame divided into blocks, compare block difference.

Compression Methods  Motion Compensation Motion vector, compare blocks with location difference. Motion vector, compare blocks with location difference.  Frame Segmentation Frame divided into equal-size nonoverlapping blocks, block size affecting the chance of matching Frame divided into equal-size nonoverlapping blocks, block size affecting the chance of matching  Search Threshold Compare #difference with a predefined value Compare #difference with a predefined value

Compression Methods  Block Search Search area = (width + 2dx)(height + 2dy) Search area = (width + 2dx)(height + 2dy)  Distortion Measure Use Pel difference instead of mean absolute difference Use Pel difference instead of mean absolute difference

Compression Methods  Suboptimal Search Methods Do some search in the search area, not all. Do some search in the search area, not all.  Coding Motion Vector Use prediction motion vector Use prediction motion vector Grouping motion vectors Grouping motion vectors  Image Objects Use overlapping object layers Use overlapping object layers

My Ideas  “Cutting-Sandwich Compression”  Motivation: for high quality compression, or lossless video compression  compression rate is low.

Cutting-Sandwich Compression  Cutting Frames horizontally or vertically, then do compression  Reason: for a single frame, the repeating rate of pixels is low  Assumption: the repeating chance of the same line of frames may be higher.

Implementation  Base on the general Video compression approach, treat images in another dimensions (horizontal cutting images, or vertical cutting images)  Treat as a 3-d block of frames, stored in buffer for encoding / decoding.

Implementation  For each 3-d block, compare the general, horizontal and vertical compression rate, take the best.  Add two extra bits for mode selection. 00 for general mode, 01 for vertical mode, 10 for horizontal mode and 11 for open. 00 for general mode, 01 for vertical mode, 10 for horizontal mode and 11 for open.  Tradeoff is low.

Tradeoff  For example, 1 block = 640 frames of (640 x 480) image, 3 bits per pixel.  2 extra bits / (640 x 640 x 480 x 3) bits ~= 0%.

Tradeoff  More resources are needed for the compression (time, software, hardware)  Encoding rate may be slower, Decoding rate – no big difference if we take some buffer (again, only decoding needed to be fast in general case)  Software and hardware are cheap now, not a technical problem.

Tradeoff  Buffer is needed for encoding and decoding in video compression, not a big deal.

Other supportive compression methods in Sandwich  Run-Length Encoding  LZ77  LZW