Visual Intelligence

Imaging of the unseen is needed in a large variety of applications: imaging the inside of the body, the seawaters, the subsea and the earth from above. However, there is a lack of analysis tools with the power and trustworthiness to fully exploit such complex imagery for knowledge extraction, innovations, and reliable new technology.

Vision

Visual Intelligence shall be the lead provider of cutting-edge solutions for complex image analysis by leveraging deep learning to answer innovation needs shared across a consortium of corporate and public user partners from different business areas. They all rely on complex image data for sustainable value creation, posing shared research challenges. This enables crucial cross-fertilization in the research and innovation.

Knowledge needs

A big driver in the recent progress of AI systems is the use of deep learning. The largest success has been within computer vision, where the use of deep learning has led to a giant leap in performance for image object detection and recognition. The progress has been immense in general computer vision applications for images taken from handheld optical cameras, facilitated by vast datasets collected by actors like Google or Facebook. This has led to a range of new image-based technologies that is rapidly changing society. Despite these advances, it is still a long way before the potential of deep learning for visual intelligence is realized for applications and industries relying on more complex visual data.

An illustration of the relationship between the research challenges and innovation areas of Visual Intelligence.

The innovation gap between current state-of-the-art and its potential is particularly large for user applications where the amount of annotated visual data is limited, and experts are needed to interpret the data. The Visual Intelligence research partners, UiT – The Arctic University of Norway, Norsk Regnesentral (NR, Norwegian Computing Center) and the University of Oslo (UiO), will answer these challenges. Creating synergies and cross-fertilization between different applications that all depend on complex visual data to enable new deep learning methodology, we will contribute to realize this potential for four core selected innovation areas of great societal importance:

  • medicine and health
  • marine science
  • energy and industry
  • earth observation.

Visual Intelligence includes the following major private and public partners: Equinor, GE Vingmed Ultrasound, the Cancer Registry, University Hospital of North Norway (UNN), Helse Nord IKT, Institute of Marine Research (IMR), Kongsberg Satellite Services (KSAT), and Terratec. Together we will enable visual intelligence for extracting crucial and actionable knowledge from complex visual data. Visual Intelligence will focus on deep learning research challenges shared by the different application domains, enabling translation of solutions between domains.

Our innovations will rely on development of explainable deep learning models, utilizing limited training data and geometrically and physically plausible decisions with associated reliability and uncertainty, solving the grad research challenges across all domains.

Objectives

Our main objective is to unlock the potential of visual intelligence across the main innovation areas medicine and health, marine science, energy sector, and earth observation by enabling the next generation deep learning methodology for extracting knowledge from complex image data.

The secondary objectives aim at this through analysis of complex imagery for real-life applications with:

  • Solutions for learning from limited data
  • Solutions for exploitation of context, dependencies, and prior knowledge
  • Solutions for estimation of confidence and quantification of uncertainties
  • Solutions for explainable and reliable predictions

The innovations will result in new and improved products and services for value creation for the user partners.

To achieve these goals, Visual Intelligence work across the innovation areas to extract information from large volumes of visual data more efficiently through automatic and intelligent data analysis. The core challenges in deep learning: working with limited training data, utilizing context and dependencies, providing explainability, confidence and uncertainty, are important in all the innovation areas.

A common challenge across the innovation areas is the need for efficient processing and analysis of large amounts of visual data. There will be a significant amount of synergy between the innovation areas. Only parts of the data are annotated and can be used for supervised learning, hence, the innovation areas share the need for utilizing semi-supervised and weakly/unsupervised deep learning. Furthermore, the areas share a need for incorporation of context, dependencies, and prior knowledge, as these factors are often crucial for obtaining the needed performance.

Medical x-ray image.

As the applications involve imaging the unseen – the inside of the human body, the sea, the subsurface, and the surface of the earth seen from space independent of daylight and weather conditions, the need for explainable results with associated uncertainties and confidence levels are particularly important as the human eye cannot provide a straightforward solution.

The innovation areas are applications of importance to the Norwegian society, and it is imperative that Norway is able to build deep learning models trained on Norwegian data. The partners in Visual Intelligence have decades of experience in working with these types of data, and this domain-knowledge will be crucial.

Supporting the UN sustainability goals

The results and the innovations aimed at through Visual Intelligence can contribute to several challenges and important issues addressed by the UN’s sustainability goals.

The innovations in the field of medical image analysis contribute directly to ensuring healthy lives and as well as promoting wellbeing for all at all ages (goal 3) by obtaining more efficient tools for diagnosing heart disease and cancer.

Our solutions for monitoring the marine environment and fisheries contribute to sustainable use of the oceans and marine resources (goal 14) while innovations for vegetation monitoring contribute to sustainable management of forests, land resources and biodiversity (goal 15).

Improved solutions for interpreting the digital subsurface can greatly help ensure access to affordable, reliable, sustainable and modern energy (goal 7).

UN sustainability wheel.