.PDF File Format – Images

Image XObjects

The data, after the filters are undone, is the samples row by row from the top left, each row padded to a whole byte, components interleaved: Height × ⌈Width × components × BitsPerComponent / 8⌉ bytes. Image-format filters (DCT, JPX, JBIG2, CCITT) produce samples in that layout themselves.

Colour spaces and bit depths

An image may use any colour space except Pattern. Indexed images may have 1, 2, 4 or 8 bits a sample; 16 bits, from PDF 1.5, is for the other spaces. Typical forms:

Decode arrays

Decode has two numbers, Dmin and Dmax, for each component. A sample x of n bits maps to Dmin + x × (Dmax − Dmin) / (2ⁿ − 1). The defaults:

Swapping the pair inverts a component: [1 0] on a grey image makes a negative, and on a stencil mask paints the 1 bits instead of the 0 bits.

Masks

Four ways to show only part of an image:

• Stencil mask (1.0): ImageMask true, one bit a pixel, no colour space. The 0 bits are painted in the current fill colour and the 1 bits left alone (with the default Decode [0 1]). Used for one-colour scanned text and Type 3 bitmap glyphs.
• Explicit mask (1.3): Mask is another image, a stencil mask of any resolution, laid over the same unit square. Where it is 1, the image is hidden.
• Colour-key mask (1.3): Mask is an array of a minimum and maximum for each component, such as [250 255 250 255 250 255]. Pixels whose every component falls in the ranges are not painted, as with a transparent colour in GIF.
• Soft mask (1.4): an image of opacity, for smooth edges and partial transparency.

Soft masks

An image’s SMask, from PDF 1.4, is an image XObject of its own, in DeviceGray, whose samples give the opacity of the image at each point: 0 transparent, 1 opaque. It need not have the same size as the image; both are mapped onto the same unit square. Its keys are those of an image, without Mask, SMask or ImageMask, and with one of its own:

A PNG with alpha becomes, in a PDF, an RGB image and a grey soft mask, each Flate compressed. When SMask is present, Mask is ignored; a soft mask set in the graphics state is overridden for the image as well.

Image data

• DCTDecode – a whole JPEG file. Most viewers hand it to a JPEG decoder as it is, which is why a PDF can hold a JPEG that can be extracted byte for byte
• JPXDecode – a JPEG 2000 file or codestream, from PDF 1.5; see below
• JBIG2Decode – a one-bit JBIG2 image, with its global segments in a separate stream
• CCITTFaxDecode – a one-bit fax-coded image, as in a TIFF Group 4 file
• FlateDecode, LZWDecode, RunLengthDecode – raw samples, compressed without loss

Only the image filters change how the samples are laid out. With every other filter the decoded data must already be the rows of samples described above.

JPEG 2000 images

A JPXDecode image, from PDF 1.5, can carry what the image dictionary would otherwise say. The rules:

• ColorSpace may be left out. The colour space is then the one in the JPEG 2000 data: a JP2 colr box with an enumerated space or an ICC profile. If ColorSpace is given, it overrides the data’s.
• BitsPerComponent is ignored; the data says. Components may differ in depth.
• Decode is ignored, except for an image mask.
• SMaskInData says what to do with an alpha channel in the data: 0 ignore it (default), 1 use it as a soft mask, 2 use it as a premultiplied soft mask. If SMask is given, it wins.
• A JPEG 2000 image may be an image mask, with one component of 1 bit.

PDF/A-1, based on PDF 1.4, does not allow JPEG 2000; PDF/A-2 and later allow it with limits on its colour boxes and bit depths.

Inline images

A small image can be written straight into the content stream:

q 17 0 0 17 298 388 cm
BI /W 17 /H 17 /CS /RGB /BPC 8 /F [/AHx /Fl]
ID
789CF3… >
EI Q

Between BI and ID come key and value pairs, usually in short forms. After ID comes one white-space byte, then the data, then EI.

Values may be short too: colour spaces G, RGB, CMYK and I (Indexed), and the filter names AHx, A85, LZW, Fl, RL, CCF and DCT. A colour space may also be a name in the ColorSpace resources, and an Indexed space may be written in full with its lookup string. There are no short names for JBIG2Decode and JPXDecode, and an inline image has no way to refer to another stream, so it has no soft mask, no mask image and no JBIG2 globals. The standard advises keeping inline images to 4 KB or less, and making larger ones XObjects.

Finding the end is the hard part, since the data may hold the bytes EI. Readers compute the data’s length from the width, height and depth when there is no filter, decode the filter to find where it stops when there is one, or look for EI preceded by white space and followed by white space or the end. PDF 2.0’s L settles it where writers give it.

Alternates

Alternates, from PDF 1.3, is an array of alternate image dictionaries: each has Image, another image XObject, and DefaultForPrinting, true for the one to print. A viewer may show a low-resolution image on screen and print a high-resolution one. Few readers use it, and PDF/A forbids it.