Design Criteria for General-Purpose Still-Image Compression


Binary Tree Predictive Coding is a solution to the problem of general-purpose still-image compression. The criteria it was designed to meet are as follows:

  1. The scheme should support lossless and lossy compression, with no fundamental change in the algorithm.
  2. The rate-distortion performance of the scheme for lossy compression of continuous images should be comparable to that of JPEG. In tests on BTPC, "comparable" was quantified as nowhere lower than 2dB reconstruction PSNR below JPEG, and on average as good as JPEG. [In practice, BTPC 2 is usually superior at the compression ratio achieved by JPEG when its quality parameter is set at 75, the usual default.]
  3. The lossless compression performance for continuous (photographic) images should be better than GIF, and for graphical images should be within 1 bpp (bit per pel) of GIF.
  4. Lossy compression of graphical images, for example lettering, captions on photographs, and purely geometric forms, should be possible, monotonic and visually consistent. That is, the picture PSNR should degrade monotonically as the data rate is reduced, and this degradation should visually appear as a gradual worsening of the picture (by blurring) rather than as sudden changes in quality. Presentation graphics, combining photographic-type shading and texturing with the traditional hard lines of graphics, should be monotonically and efficiently encoded. For graphical images which losslessly compress more under GIF than under the general-purpose scheme, the visual degradations due to lossy compression at the data rate achieved by GIF should be invisible.
  5. Software implementation of the decoder should be memory and time efficient. The decoder should require little more memory space than the size of the image. The number of operations required to decode an image should be a small multiple of the image size.
  6. Software implementation of the encoder should also be efficient; however, greater complexity is allowed because encoding is relatively infrequent.
  7. Hardware implementation of the encoder and decoder should allow for fine-grain paralelism, ideally down to one processor per pel.

Next (Image Pyramids and Trees) | Top (BTPC)