.JPG File Format – Frame header

JPEG image (JFIF and Exif)

The frame header says how large the image is, how many components it has and how they are sampled. Its marker also says how the image was coded. It is defined in section B.2.2 of ITU-T T.81.

Coding processes

The marker's code picks the process. Most programs handle only the first three. SOF4, SOF8 and SOF12 do not exist: those codes are DHT, JPG and DAC.

MarkerNameProcessSample bits
FF C0SOF0Baseline sequential DCT, Huffman coding8
FF C1SOF1Extended sequential DCT, Huffman coding8, 12
FF C2SOF2Progressive DCT, Huffman coding8, 12
FF C3SOF3Lossless sequential, Huffman coding2 to 16
FF C5SOF5Differential sequential DCT, Huffman coding8, 12
FF C6SOF6Differential progressive DCT, Huffman coding8, 12
FF C7SOF7Differential lossless, Huffman coding2 to 16
FF C9SOF9Extended sequential DCT, arithmetic coding8, 12
FF CASOF10Progressive DCT, arithmetic coding8, 12
FF CBSOF11Lossless sequential, arithmetic coding2 to 16
FF CDSOF13Differential sequential DCT, arithmetic coding8, 12
FF CESOF14Differential progressive DCT, arithmetic coding8, 12
FF CFSOF15Differential lossless, arithmetic coding2 to 16

Baseline uses 8-bit samples only. The processes marked differential belong to the hierarchical mode, in which several frames follow a DHP segment.

The segment

  1. 0Marker FF C0 or another SOFn2 bytes
  2. 2Lf segment length, 8 + 3 × Nf2 bytes
  3. 4P sample precision in bits1 byte
  4. 5Y number of lines2 bytes
  5. 7X samples per line2 bytes
  6. 9Nf number of components1 byte
  7. 10Component 13 bytes
  8. …Further components3 bytes each
A frame header (offsets from the marker)
OffsetSizeFieldDescription
41PBits per sample
52YImage height. 0 means a DNL segment after the first scan gives the height.
72XImage width, at least 1
91NfNumber of components: 1 to 4 in a progressive image, up to 255 in sequential and lossless

Each component

OffsetSizeFieldDescription
01CiComponent identifier. Scans name components by this.
1½Hi (high 4 bits)Horizontal sampling factor, 1 to 4
1½Vi (low 4 bits)Vertical sampling factor, 1 to 4
21TqiNumber of the quantisation table the component uses, 0 to 3

JFIF files use 1 for Y, 2 for Cb and 3 for Cr. A file with components named by the letters R, G and B (82, 71 and 66) is usually taken to be RGB, with no colour transform.

Sampling and blocks

Components can be stored at different resolutions. Hmax and Vmax are the largest factors in the frame. Component i has x = ⌈X × Hi / Hmax⌉ samples across and y = ⌈Y × Vi / Vmax⌉ lines down.

The samples are divided into blocks of 8 × 8, called data units. A component coded by itself has ⌈x / 8⌉ × ⌈y / 8⌉ blocks. A component coded with others in the same scan follows the MCU layout below, which rounds up to whole MCUs, so its blocks can be more.

  1. YY blocks Hi × Vi of them, in rows, left to rightHi × Vi blocks
  2. CbCb blocksHi × Vi blocks
  3. CrCr blocksHi × Vi blocks
One MCU of a colour scan. In a 4:2:0 image Y is 2 × 2, and Cb and Cr 1 × 1, so there are six blocks.

Minimum coded units

An MCU is the group of blocks coded together. In a scan with one component it is one block. In a scan with several components it covers 8 × Hmax pixels by 8 × Vmax, and holds Hi × Vi blocks of each component, one component after another, each component's blocks left to right and top to bottom. An MCU may hold at most 10 blocks. MCUs at the right and bottom edges are always whole: a component that does not fill one is padded out with blocks of its own, which a decoder discards.

The number of MCUs is ⌈X / (8 × Hmax)⌉ across and ⌈Y / (8 × Vmax)⌉ down.

Common nameYCb, CrMCU sizeBlocks in an MCU
4:4:41 × 11 × 18 × 83
4:2:22 × 11 × 116 × 84
4:2:02 × 21 × 116 × 166
4:1:14 × 11 × 132 × 86

DNL: the height, sent later

When Y is 0, a DNL segment (FF DC, length 4, 2-byte line count) directly after the first scan gives the height. Such files are rare, and browsers do not read them.