Image compression and active acquisition using deep posterior sampling

Researcher:
Prof. Michael Elad | Computer Science
Prof. Tomer Michaeli | Electrical and Computer Engineering

Categories:

Information and Computer Science | Security & Defense

The Technology

Signals compression or active acquisition can be very heavy data transmitting processes.
By taking advantage of recently discovered generative AI capabilities (and specifically diffusion models), our technology enables a computational process for discovering an optimized set of measurements for obtaining maximal information on a signal of interest. This task is known as “compressive sensing”, and our solution offers a signal adaptive sequential process for sensing the signal.
Our platform may be applied in the following fields for example:

  1. This can be beneficial in medical imaging, such as CT or MRI, where multiple sensing operations are applied jointly to reconstruct a visualization used in diagnostics. Our method can greatly reduce the number of measurements required to produce the same visualization, increasing the speed of the medical imaging acquisition, and even reducing patient exposure to ionizing radiation, in the case of CT.
  2. Using the same rationale, our technology can be turned into a highly effective lossy compression algorithm. As our method discovers the most meaningful measurements, those can be saved or transmitted as a compressed representation of an original image or signal. The decoder, running the same generative AI method, does not need to get any side information beyond the quantized measurements, and this leads to a progressive compression scheme that uses an effective data-adaptive transform.

Advantages

  • No need for task specific training or tuning. Just a single foundation generative model
  • Superior to other classical approaches. High fidelity that ensures results realism and faithfulness to the original data

Applications and Opportunities

  • Versatile platform that can be applied to active acquisition, compression and many more, without additional data or computation
  • Compression can be applied to images, video, audio and other signals, for which effective coding schemes are non-existent
  • Patent pending technology
  • A team of researchers experts in signal processing and generative AI
arrow Business Development Contacts
Dr. Arkadiy Morgenshtein
Director of Business Development, ICT