Overview

The Adobe DNG standard is the only widely supported open way to exchange raw camera sensor data between various pieces of software. It uses the TIFF format to store the sensor data and has a series of custom TIFF tags to describe the method to convert the raw sensor data to regular RGB color data.

DNG Limitations

In order to produce valid dng files from sensor data the only valid options are a grayscale image produced by a camera with a color filter array (a bayer filter) or raw data which has been demosaiced into a multi-channel color image.

It is not possible to store YUV encoded image data or any of the UVC camera supported data formats.

Bit packing of sensor data

Most camera sensors that produce raw data connected to a Linux system will be connected over the MIPI-CSI bus. The MIPI specifications declare several formats for transmitting the pixel data. For modes where the camera returns 8-bit data this is exactly what you expect with one byte mapping to one sample. For 16-bit data it simply maps to two bytes per sample but a lot of sensor data will be either 10 or 12 bits per sample.

The packing in the MIPI specification is matching to the packed formats in v4l2: V4L2_PIX_FMT_SBGGR10P and V4L2_PIX_FMT_SBGGR12P and their other 4 corresponding CFA orders. These formats pack the 8 most significant bits of the 4 samples as seperate bytes and then 1 or 2 extra bytes that stores all the “leftover” bits.

In the diagram above the first samples from a camera in BGGR mode is shown. The first line has alternating blue and red color filters here shown in two shades of blue and green to differentiate the 4 samples. The semi-transparent bits are the least significant bits of each sample.

The DNG standard specifies a subset of TIFF image data encodings for storing the sensor data into the DNG. These formats sadly do not line up with the pixel format of MIPI sensor data like most cameras use. The libdng code provides some helper code to convert the 10 and 12-bit MIPI pixel formats to the more linear format used in the TIFF file format.

Writing a simple DNG file

The minimum code for writing a valid DNG file is the following:

#include <libdng.h>

// Imagine you have real data
uint8_t *data = <from somewhere>;
size_t data_length = 1234;

// Run once to register custom tags in libtiff
libdng_init();

// New image
libdng_info dng = {0};
libdng_new(&dng);

// Set the pixel format to 10-bit RGGB in MIPI format
if (!libdng_set_mode_from_name(&dng, "SRGGB10P")) {
  fprintf(stderr, "Invalid pixel format supplied\n");
}

// Call any other libdng_set_* methods here

if (!libdng_write(&dng, "out.dng", width, height, data, data_length)) {
  fprintf(stderr, "Could not write DNG\n");
}

libdng_free(&dng);

This writes out the raw sensor data from *data to the out.dng file with the right metadata for a RGGB color filter array. It also converts the 10-bit packed format to the linear format from the TIFF spec.

Operations that change metadata can be executed in any order between the libdng_new and libdng_write command. The ony required metadata is one of the libdng_set_mode_* calls to bring all the pixelformat metadata in order.

Setting the mode

The only required metadata operation is setting the pixelformat. There are multiple ways of specifying the format:

• With libdng_set_mode_from_name(&dng, "name"):

  This is setting the format from one of the named modes. The accepted named modes here are the 4 character fourcc codes defined by V4L2 and also the more intuitively named V4L2 constants without the V4L2_PIX_FMT_ prefix in front.

• With libdng_set_mode_from_pixfmt(&dng, pixfmt):

  This does the same thing but instead of passing the format as string you pass the value of the V4L2_PIX_FMT_ constants as an integer. This is easier for integrating the DNG writing with existing V4L2 code.

If the format is one of the packed formats the libdng_write method will automatically convert the format to fit in a TIFF container.