Deep Red is a chair of excellence supported by the Grenoble INP Foundation in partnership with LYNRED, a world leader in the development and production of high quality infrared technologies. This chair was created with the aim of exploiting the full potential of infrared sensors by taking advantage of the latest advances in data science, artificial intelligence and deep learning.

Indeed, a raw infrared image contains a high level of noise that can hinder its optimal exploitation. In order to improve the readability of imaged scenes, whether by a human operator or by a processing algorithm aiming, for example, at detecting anomalies or objects, different strategies can be explored.

One approach consists in proposing g methods of fusion of visible and thermal infrared images , intended in particular for augmented reality and automotive applications. We will then seek to take advantage of the complementarity of visible and infrared imaging to increase the image quality from the point of view of human or machine vision. This complementarity would translate on the merged image by an improvement of the resolution, a reduction of the noise, an improvement of the colorimetry, of the contrast, and this, according to various climatic conditions and types of scenes.

Another way is to usedeep learningg models for noise reduction, information fusion or object detection. In order to set up sufficiently frugal architectures, possibly embedded AI, the integration of expert knowledge on infrared imagery and the corresponding sensors will be crucial. We will talk about physics-driven artificial intelligence. This approach, which aims to learn from data what modeling cannot estimate analytically, will come as an alternative to traditional analysis and processing methods with the objective of "getting the best of both worlds".embedded AI), l’intégration de connaissances expertes sur l’imagerie infrarouge et les capteurs correspondants seront capitales. On parlera d’intelligence artificielle guidée par la physique. Cette approche, qui vise à apprendre des données ce que la modélisation ne sait pas estimer de manière analytique, viendra comme une alternative aux méthodes traditionnelles d’analyse et de traitement avec comme objectif de « tirer le meilleur des deux mondes ».