History of the reverse engineering of the st2205 compression ------------------------------------------------------------ After some failed attempts by me (Hans de Goede), I asked help from Bertrik Sikken with reverse engineering the st2205 compression. He went the route of disassembling the windows binaries for uploading pictures, and wrote an algorithm description of the decompression on the picframe wiki: http://picframe.spritesserver.nl/wiki/index.php/ImageEncoding A copy of this page is included below. I used this description to write decompression code. The first version used header files with the lookup tables as C-code arrays and these header files were made by and given to me by Bertrik. I also wrote compression code by "simply" reversing the decompression algorithm. The lookup tables were a problem, as Bertrik copied them out of the windows code, and one could argue that they are copyrightable. Luckily the same tables are also present in the firmware of the picture frame, and we can read them from the frame. So the libgphoto2 st2205 camlib uses this approach to avoid any copyright issues. Below is a dump of: http://picframe.spritesserver.nl/wiki/index.php/ImageEncoding Made on 19 March 2010. Contents • 1 Compressed image format □ 1.1 Introduction □ 1.2 Image decoding ☆ 1.2.1 Image header ☆ 1.2.2 8x8 pixel block ○ 1.2.2.1 Decoding chrominance ○ 1.2.2.2 Decoding luminance ○ 1.2.2.3 Color decoding ☆ 1.2.3 Shuffling/unshuffling □ 1.3 Table overview Compressed image format Images written on the keychain by the Photoviewer application are stored in a compressed form. This section describes the compressed image format. NOTE: code for reading images from a ST2205-based picture photoframe has been included now (march 2010) in gphoto SVN Introduction The photo keychain can contain several images. A simple directory-like structure starting at address 0 contains pointers to where the actual compressed images are stored. Basic principles used in the encoding: • 8x8 image blocks: an image has a dimension of 128x128 pixels and is subdivided into 256 blocks of 8x8 pixels each. • YUV-like encoding: each of these 8x8 pixel blocks is separated into a brightness component and two color components and independently encoded by a lossy compression algorithm. • block shuffling: the order in which the blocks are encoded is usually not simply linear top-down left-right, but is controlled by one of several "shuffling tables" that determine where each decoded block goes into the final image. By showing the decoding process in real time and using a different shuffling table for each picture, the photo key chain can create various transition effects between the previously decoded photo and the currently decoding photo. Note the above (and below) 128x128 resolution and thus 256 blocks assumes a frame with a lcd a resolution of 128x128, for displays with a different reslution this numbers change accordingly. Image decoding An image is encoded as a 16-byte image header, followed by the 256 blocks encoding each for an area of 8x8 pixels. Typically, each 8x8 block is encoded using 48 bytes (or in some special cases 56 or 64 bytes, see below), this means a total size of 12304 bytes. Uncompressed, the image would require 32768 bytes (assuming 16-bit RGB565 format), saving space by a factor of about 2.66. Image header An image starts with a 16-byte header: ┌─────────┬───────────────────────────────────────────────────────────────┐ │ Byte │ Meaning │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x00 │Image marker with a fixed value of 0xF5 │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x01/0x02│Image width (0x80) encoded as big-endian │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x03/0x04│Image height (0x80) encoded as big-endian │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x05/0x06│Number of 8x8 blocks in the image (0x100) encoded as big-endian│ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x07 │Shuffle pattern to use │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x08 │Unknown, bits 1 and 2 carry some special meaning. Usually │ │ │contains value 0x04. │ ├─────────┼───────────────────────────────────────────────────────────────┤ │ │Related to block shuffling, exact meaning unknown. Contain 0xFF│ │0x09 │for shuffle pattern 1, for other shuffle patterns it contains a│ │ │value depending on the picture frame resolution. For 128x128 │ │ │frames it contains 1, for 96x64 frames it contains 0. │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x0A/0x0B│Length of following image data │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x0C-0x0F│Padding/unused (all 0x00) │ └─────────┴───────────────────────────────────────────────────────────────┘ 8x8 pixel block The 8x8 pixel block is encoded further as a variable length block, as follows: ┌──────┬───────┬──────────────────────────────────────────────────────────┐ │Offset│Meaning│ Remark │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │ │Length │Bit 0-6 is the length of the rest of the block (often │ │0x00 │byte │0x2F). Bit 7 selects between 2-bit and 4-bit luma decoding│ │ │ │mode │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │ │Luma Y │Bit 0-6 is the base offset of the luminance channel for │ │0x01 │byte │every pixel in the block. Bit 7 selects between two │ │ │ │luminance patterns tables (LUMA1 or LUMA2) │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │ │Chroma │Bit 0-6 is the base offset of the U chrominance channel │ │0x02 │U byte │for every pixel in the block. Bit 7 indicates extended │ │ │ │decoding of this chrominance channel │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │ │Chroma │Bit 0-6 is the base offset of the V chrominance channel │ │0x03 │V byte │for every pixel in the block. Bit 7 indicates extended │ │ │ │decoding of this chrominance channel │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │0x04 │Chroma │Variable length chrominance channel U decoding info, see │ │ │U data │below │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │... │Chroma │Variable length chrominance channel V decoding info, see │ │ │V data │below │ ├──────┼───────┼──────────────────────────────────────────────────────────┤ │... │Luma Y │Variable length luminance channel decoding info, see below│ │ │data │ │ └──────┴───────┴──────────────────────────────────────────────────────────┘ Decoding chrominance Chrominance is encoded with a reduced resolution of 4x4 "pixels" per block. The chrominance decoding info consists of: ┌─────────┬──────────┬────────────────────────────────────────────────────┐ │ Offset │ Meaning │ Remark │ ├─────────┼──────────┼────────────────────────────────────────────────────┤ │ │Pattern │Index into a color base pattern table (CHROMA). The │ │0x00 │byte A │table entry contains the pattern for the top 8 │ │ │ │chroma pixels. │ ├─────────┼──────────┼────────────────────────────────────────────────────┤ │ │Pattern │Index into a color base pattern table (CHROMA). The │ │0x01 │byte B │table entry contains the pattern for the bottom 8 │ │ │ │chroma pixels. │ ├─────────┼──────────┼────────────────────────────────────────────────────┤ │ │ │Optional correction items for each chroma pixel │ │0x02-0x09│Correction│(using table CORR). This part is only present when │ │ │ │bit7 of the chroma byte is set. │ └─────────┴──────────┴────────────────────────────────────────────────────┘ The chrominance value is basically the sum of: 1. chrominance base value (from the chroma U or V byte minus 0x40) 2. one of the two base patterns selected from table CHROMA using pattern byte A and B. 3. optionally a set of correction values selected from table CORR (very much like luminance) Chroma pixels in a segment are arranged in the following pattern: A00 A01 A02 A03 A04 A05 A06 A07 B00 B01 B02 B03 B04 B05 B06 B07 where Axx indicates data looked up using pattern byte A, and Bxx data looked up using pattern byte B. Note that the chroma info per channel (U and V) is either 2 or 10 bytes, which is why the total block size can be one of 48, 56, or 64 bytes (resp no corr, corr for 1 channel, corr for both channels). By far most blocks are 48 bytes big. Decoding luminance The luminance decoding info consists of: ┌─────────┬───────────────────────────────────────────────────────────────┐ │ Offset │ Meaning │ ├─────────┼───────────────────────────────────────────────────────────────┤ │0x00-0x07│A set of indexes into table LUMA1 or LUMA2 to select a "base │ │ │pattern" for each row of pixels inside a block. │ ├─────────┼───────────────────────────────────────────────────────────────┤ │ │Correction items for each pixel in the block, where each │ │0x08-0x27│correction item is encoded in 4 bits. Each correction item is │ │ │actually an index into table CORR storing a value that is added│ │ │to the luminance value of the pixel. │ └─────────┴───────────────────────────────────────────────────────────────┘ In short, the luminance channel of each pixel is basically the sum of: 1. luminance base value as encoded in byte 0x01 2. one of the base patterns selected from table LUMA1 or LUMA2 using bytes 0x00-0x07 (one byte for one pattern per row of 8 pixels) 3. one of the correction values selected from table CORR using the nibbles in bytes 0x08-0x27. This uses one nibble for each pixel, with the high 4 bits (high nibble) of a byte coding the correction for the first pixel that bytes corrects, and the low 4 bits code the correction for the second pixel. The first byte (at offset 0x08), codes the correction for pixel 00 and 01, etc. Pixels are arranged in the following pattern: 00 01 02 03 04 05 06 07 08 09 0A 0B 0C 0D 0E 0F ... 38 39 3A 3B 3C 3D 3E 3F Color decoding The color space used in the images is a kind of YUV, using one luminance channel (called Y here) and two chrominance channels (called U and V here), as follows: R = 2 * (Y + V) G = 2 * (Y - U - V) B = 2 * (Y + U) Any underflows/overflows during the calculation of the RGB values should be saturated to either 0 or 255 respectively. And conversely (during encode): Y = (R + G + B) / 6 U = B/2 - Y V = R/2 - Y where U and V should be constrained to the range -64 to +63. Shuffling/unshuffling Blocks of 8x8 pixels are stored in a shuffled sequence. The choice of the shuffle pattern is stored in byte 0x07 of the image header. The shuffle pattern is stored in table SHUFFLEx, which contains the (x,y) coordinates of each 8x8 pixel block. Currently there are 7 known shuffle patterns. Table overview Several tables are involved in encoding/decoding. Below is an overview of the various tables, their properties and what they are used for ┌────────┬────────────┬───────────────────────────────────────────────────┐ │ Table │ Dimension │ Remark │ ├────────┼────────────┼───────────────────────────────────────────────────┤ │ │256 entries │Entry contains brightness patterns for 1 row of 8 │ │LUMA1 │of 8 signed │pixels within an 8x8 block. │ │ │words │ │ ├────────┼────────────┼───────────────────────────────────────────────────┤ │ │256 entries │Entry contains brightness patterns for 1 row of 8 │ │LUMA2 │of 8 signed │pixels within an 8x8 block. Alternate table. │ │ │words │ │ ├────────┼────────────┼───────────────────────────────────────────────────┤ │ │256 entries │Entry contains chroma patterns for 2 rows of 4 │ │CHROMA │of 8 signed │pixels within an 4x4 chroma block. │ │ │words │ │ ├────────┼────────────┼───────────────────────────────────────────────────┤ │ │256 entries │Entry contains the (x,y) coordinates for an 8x8 │ │SHUFFLEx│of 2 bytes │block. There are several tables like this, each one│ │ │ │encoding for a different shuffle pattern. │ ├────────┼────────────┼───────────────────────────────────────────────────┤ │ │16 entries │Correction values that are applied for each pixel. │ │CORR │of 1 signed │Values in this table: │ │ │word │-26,-22,-18,-14,-11,-7,-4,-1,1,4,7,11,14,18,22,26 │ └────────┴────────────┴───────────────────────────────────────────────────┘ The LUMA1, LUMA2 and CHROMA tables are stored inside the picframe in 16 bits signed (2's complement) LE byte order format. LUMA1 starts at 0x8477, LUMA2 at 0x9477, CHROMA at 0xA477. These 3 tables are directly followed at 0xB477 by various shuffle tables, which contain x, y coordinate pairs (1 byte for each). There are a number of sets of shuffle tables, for different display resolutions. There are 8 or 7 tables per set, the first 2 tables are "generated" (0 = row by row, 1 = column by column), then 6 or 5 tables per resolution in ROM: ┌──────────┬──────┬─────────────────────┐ │Resolution│start │Number of sets in ROM│ ├──────────┼──────┼─────────────────────┤ │128 x 160 │0xB477│6 │ ├──────────┼──────┼─────────────────────┤ │128 x 128 │0xC377│5 │ ├──────────┼──────┼─────────────────────┤ │120 x 160 │0xCD77│5 │ ├──────────┼──────┼─────────────────────┤ │96 x 64 │0xD92F│5 │ └──────────┴──────┴─────────────────────┘ Retrieved from "http://picframe.spritesserver.nl/wiki/index.php/ ImageEncoding" Powered by MediaWiki GNU Free Documentation License 1.2 • This page was last modified 20:49, 16 March 2010. • This page has been accessed 4,256 times. • Content is available under GNU Free Documentation License 1.2.