Bringing physics into image compression
● Validated and supported by
European Space Agency
University of Geneva
Viventis Microscopy
University of Glasgow
PCO Imaging
❝The exponential growth in the volumes of accessible imaging data across the multitude of commercial applications in conjunction with the prolifiration of AI-based analytics calls for a radically new approach for data managements and compression. The AI-optimised image compression by Dotphoton allows for significant optimisation in speed and performance of our image analytics pipeline and offers even greater potential for the compression of multi-spectral and hyperspectal airborne and space-borne imaging data.❞
Yosef Akhtman, CEO and co-founder Gamaya
● Performance. Data based on test of 1024 images. See details
● Our aim is not to provide a fixed compression ratio, but to achieve the highest possible lossless compression ratio for a particular image from a particular image sensor and meanwhile guarantee full image information retention. The compression ratio varies across applications and industries, but Dotphoton was proven to a an extremely performant algorithm.
● Our aim is not to provide a fixed compression ratio, but to achieve the highest possible lossless compression ratio for a particular image from a particular image sensor and meanwhile guarantee full image information retention. The compression ratio varies across applications and industries, but Dotphoton was proven to a an extremely performant algorithm.
● It is the purpose of image sensors to measure the amount of light that is emitted by an object in a given direction or at a given position. Under normal circumstances, the physical processes behind the emission are somewhat random or irregular. The irregularity manifests itself as noise in the output image. Unintuitively, the better the image sensor, the more appearant is this fundamental noise, so for modern cameras it dominates the noise of the device itself, except for extreme low-light situations. Dotphoton specialises in characterising and understanding image sensors and the physics of image acquisition. Based on this deep understanding, we improve image processing tasks.
Compression, for example, is fundamentally limited by the presence of noise.
This is the reason why artefacts appear in most types of lossy compression: the algorithms try to remove the noise, for example by smoothening, to achieve a smaller output. A much more reliable approach, developed by Dotphoton, is to replace the noise with pseudo noise. Dotphoton collaborates with device manufacturers to create a detailed noise model of the image sensor, such that the pseudo noise has the same characteristics as the real noise. It has been shown that the resulting images produce the same results in a variety of situations as the non-processed images. In addition, new possibilities for open up, such as efficient lossless compression.
● It is the purpose of image sensors to measure the amount of light that is emitted by an object in a given direction or at a given position. Under normal circumstances, the physical processes behind the emission are somewhat random or irregular. The irregularity manifests itself as noise in the output image. Unintuitively, the better the image sensor, the more appearant is this fundamental noise, so for modern cameras it dominates the noise of the device itself, except for extreme low-light situations. Dotphoton specialises in characterising and understanding image sensors and the physics of image acquisition. Based on this deep understanding, we improve image processing tasks.
Compression, for example, is fundamentally limited by the presence of noise.

This is the reason why artefacts appear in most types of lossy compression: the algorithms try to remove the noise, for example by smoothening, to achieve a smaller output. A much more reliable approach, developed by Dotphoton, is to replace the noise with pseudo noise. Dotphoton collaborates with device manufacturers to create a detailed noise model of the image sensor, such that the pseudo noise has the same characteristics as the real noise. It has been shown that the resulting images produce the same results in a variety of situations as the non-processed images. In addition, new possibilities for open up, such as efficient lossless compression.
The performance of compression is evaluated across the axes of speed and compression ratio. Higher values are better. The unit for speed is Megapixels/s. Typical uncompressed files are 16- bits per pixel, multiply megapixels/s by 2 to estimate MBytes/s throughput. The compression ratio is with respect to the real bit-depth of the data. Typically, images having a depth of 12 or 14 bits, are still stored using 16 bits per pixel on disk. If this is the case, the user can expect higher effective compression ratios that the ones shown here.


Aerospace Bayer images, extremely rich in fine detail, optimal exposure. (source: customers)

Microscopy Monochrome images, low levels of detail, often dark and delicate amplitude variations. Important signal may be burried within the noise. (source: customers).

Photography Broad variety of images taken with many different cameras, exposures, levels of sharpness, ISO range, often include dark and saturated areas, in and out of focus regions. (source: customers).

Test images Test scene designed by Dotphoton with elements of all of the above. extremely sharp focus, and optimal exposure, taken under laboratory conditions. (source: Dotphoton).
Image sensor characterisation and modelling
Accurate noise replacement
Metrological tests
1
2
3
● The three pillars of Dotphton compression
● Dotphoton is a spinoff of the GAP-Optique, the Group of Applied Optics at the University of Geneva. Our R&D team includes experts in quantum information and quantum cryptography, computer science, microscopy. We collaborate on research with the University of Geneva, Hepia Geneve, HES Valais, Univeristy of Glasgow, EPFL and others.
● Case study. Rat's brain lightsheet microscopy. Wyss Center, Hes-so and University of Geneva
Camera: Hamamatsu ORCA Flash 4
Differences between original and compressed image < 1σ
Compression ratio of 10.5:1
1-4MB
Size of a single image
Number of images taken for a scan
Size of the final stitched image
Stitching time for 1M images
Without Dotphoton
With Dotphoton
100-800kB
100k-1M
100GB-1TB
1 week
100k-1M
10GB-400GB
1 day
● Further reading
w/o DP
w DP
Let's talk about your project.
Contact Michael Desert:

core@dotphoton.com
+41 (0) 41 552 50 00
Your
images,
lighter than light