AI Research At The French CNRS Proposes A Noise-Adaptive Intelligent Programmable Meta-Imager: A Timely Approach To Task-Specific, Noise-Adaptive Sensing

Researchers from the French CNRS have come up with a Noise-Adaptive Intelligent Programmable Meta-Imager. Sensing systems are increasingly used in many aspects of our lives, including touchless human-computer interfaces, driverless vehicles, and ambiently supported health care. These systems, however, frequently lack intelligence since they have the propensity to gather all information, regardless of whether it is pertinent. This may result in invasions of privacy as well as a loss of time, effort, and computational resources while processing data.

However, measuring procedures in practical applications are invariably impacted by different types of noise. Every measurement is inherently accompanied by noise. Particularly in indoor settings where the electromagnetic signals that are transmitted must be kept modest, the signal-to-noise ratio may be poor. In order to advance the prior research, researchers from French CNRS have now developed an intelligent programmable computational meta-imager that not only adapts its illumination pattern to a particular information-extraction task, such as object recognition, but also to various types and levels of noise.

The noise of some kind and intensity inevitably taints measurement processes. We postulate that the kind and amount of noise will affect the best coherent illumination patterns that a smart, programmable meta-imager should use to effectively extract task-specific information from a picture. It is considered a single-transmitter, single-detector multi-shot programmable computational imaging system. These systems are especially relevant in the microwave domain, where expensive transceivers can be replaced by programmable metasurface apertures, which can synthesize coherent wavefronts from a single radiofrequency chain.

The influence of latency restrictions and noise on intelligent multi-shot programmable meta-imagers is carefully explored in this article, according to the researchers. The researchers studied a common object-recognition problem and suggested a microwave computational programmable meta-imager system for it in order to test their theory. These systems might be used for earth observation, indoor surveillance, etc.

In their model, a microwave dynamic metasurface antenna (DMA) used a single transmitter to send a series of coherent wavefronts to the scene, while a second DMA used a single detector to coherently collect the reflected waves. A differentiable end-to-end information-flow pipeline was developed, consisting of the future digital processing stages as well as the programmable physical measuring process with noise.

This joint optimization, which involves task-specific end-to-end joint optimization of the trainable physical parameters and trainable digital parameters, gives the measurement process task awareness, enabling it to distinguish between information in the analog domain that is relevant to the task at hand and information that is not.

When the amount of information that can be extracted from a scene is constrained by latency constraints and/or noise, the scientists found that this programmable meta-imager, which generates a sequence of task-specific and noise-specific scene illuminations, performs better than conventional compressed sensing with random configurations.

 Gains in performance were observed for both signal-independent and signal-dependent additive noise types. Despite the “black box” character of the method, the “macroscopic” aspects of the learned lighting patterns, notably their reciprocal overlaps and intensities, were found to be intuitively accessible.

According to the researchers, the shift toward a system that autonomously recognizes the kind and quantity of noise and modifies its DMA setups correspondingly without extra human input is simple.


Check out the Paper and Reference Article. All Credit For This Research Goes To Researchers on This Project. Also, don’t forget to join our Reddit page and discord channel, where we share the latest AI research news, cool AI projects, and more.

🐝 Join the Fastest Growing AI Research Newsletter Read by Researchers from Google + NVIDIA + Meta + Stanford + MIT + Microsoft and many others...