Dear all,
I am looking into libJPEG and other libraries to do 8 by 8 pixel compression of images in a sort of tile by tiles fashion. However most of these software libraries do it by scanline and I really need the per tile both encoding and decoding in a reasonable quality and with high speed.
Anyone with experience in the area?
Creating an 8 by 8 image and encoding that (for each tile) seems a bit too cumbersome.
spinningcube
Are you sure it does its thing by scanline? Wikipedia seems to indicate that it does indeed use 8x8 blocks -- this would be expected, as there are higher odds of color cohesion in an 8x8 block, rather than, say, a 1x64 scanline.
At any rate, why JPEG specifically -- Its decent in general, but its really intended for "organic" kind of scenes (like photos of real-life things) where some details aren't important (its lossy). Is that actually your subject, or are you doing something else? For example, images with a lot of repetition would probably respond better to a block or scan-section, dictionary-based approach.
What are you trying to compress?
JPEG2000 has support for tiles of arbitrary sizes that can be decoded individually.
JPEG doesn't have proper tile support, though. It encodes images in blocks, which are already 8x8 in size (though subsampling will give a different MCU size, like 16x8 or 16x16; if you want an MCU of 8x8 then don't subsample). But these blocks aren't quite the same as JPEG2000 tiles, and if you wanted control over them you're going to have to work with a JPEG encoder/decoder on a very low level (read: you're going to have to customize the encoder/decoder).
Or maybe I misunderstood?
Thanks guys,
Might have delve a bit deeper into the codec then. I need to say block in memory ptrx to ptr y is from tile a in the image. I can myself decompose the image and assign tile indices but wanted to know if possible to force the encoder to scan certain pixels in the order that I want and not the full image in one go.
such as
image is 1024 by 768
encode starting with index pixels 8*x + width*8*y
something like that
spinningcube
One thing you could do is make your image width a multiple of 64. Then each tile is just a scanline (or half of a scanline if it's 128 wide), and it all fits in contiguous memory. Though that'll probably lead to a noisier super image that will yield less quality when compressed (which can be avoided by using a lossless codec like PNG; but the noisy super image will compress worse).
It is worth noting that jpg is generally very bad at tiled images. Fringing really screws things up and of course if you tightly pack things, the fringing overlaps into other tiles which causes problems if say you have red on the left, green in the middle and blue on the right. Your middle green tile will show artifacts of red on the left and blue on the right which can cause the tile to no longer "tile" correctly, i.e. if it were a grass texture you would likely get artifacts causing obvious seems at the edges of the tile. I would suggest Cornstalks suggestion of png also. No it does not get as good of compression but lossy formats with fringing errors can really ruin tiles.
