Couger AI team wins second place in top international AI competition and has research paper accepted by a prestigious international conference

We are delighted to share that a team from Couger Inc. (Headquarter: Tokyo, Japan), comprising of Devanathan Sabarinathan and Dr. Priya Kansal, won second place at the Thermal Image Super-Resolution Challenge, hosted by the world-renowned Institute for Electrical and Electronics Engineering (IEEE). Their research paper has also been accepted for presentation at the prestigious annual Conference on Computer Vision and Pattern Recognition (CVPR).

The Winner Certificate from the 16th CVPR Workshop on Perception Beyond the Visible Spectrum.

The team’s research involves a method for generating high-resolution images through machine learning from a pair of thermographic images obtained from different cameras.

Summary

With the explosion of data around the world, research and social implementation of AI is rapidly progressing. In the age of AI, trial and error of applied methods is carried out on an almost daily basis to realize highly accurate automation that has the ability to respond to various situations. One such area of application is the high-resolution analysis of thermographic images, which was the theme of the Thermal Image Super-Resolution Challenge.

Unlike photographic image data, thermographic images are used in a wide range of fields, including medical and military, and general object detection. Because they are not affected by ambient light or shadows, thermographic images have excellent permeability and can detect heat. These qualities can be used in various contexts. For example, identifying patients with fever at airports, which can help prevent the spread of viral infections, or for detecting people when driving at night when objects become difficult to see, which can help prevent road accidents.

Original image taken in high-resolution.

In general, thermographic images have a low resolution, so, traditionally, high-resolution images have been processed using methods such as pre-processing of data and added noise and blurring. There was almost no method of learning using a pair of images (a low-resolution image and a high-resolution image) obtained from different cameras. Such methods were one of the main focus points of the Thermal Image Super-Resolution Challenge held for new machine learning-based methods using the image-pair data.

Image generated by AI from low-resolution images.

Couger’s award-winning approach

The approach proposed in the paper (which can be read here) is based on a deep convolutional neural network, called a Multi-Level Supervision Model, using convolutional layers, residual connections, and attention modules. This technique has the following advantages:

• Because of the multi-level hierarchy, 1 model can handle high-resolution tasks at three different scales (x2, x3, x4).
• Low complexity because, by utilizing the residual connection, it focuses on the information lost.
• The robustness (*1) of this model can be improved by retaining spatial information using convolutional layers, residual learning, and CBAM (Convolutional Block Attention Module).

The architecture developed by the Couger team.

It has been proven that the Couger-proposed architecture can achieve the same level of performance not only in the dataset used in the competition but also in other datasets.

Whereas current thermographic devices are expensive and pose a financial barrier for companies wishing to introduce them, we expect to raise the level of AI implementation in our society by enabling higher resolution through machine learning even for low-resolution images acquired from inexpensive devices.

Devanathan Sabarinathan, a co-author of the paper, and Dr. Priya Kansal has also been successful in the field of computer vision, including winning 3rd place in an eye-tracking competition sponsored by Facebook and presenting a paper on skeleton estimation at the CVPR-2019 conference. Based on their, as well as other, research and development initiatives, Couger is developing applications for the technology in the development of human-like AI “Virtual Human Agents’’ that have an in-depth understanding of humans. Couger will continue to offer services and research aimed towards the social implementation of new technologies.

About IEEE

IEEE is the world’s largest academic research organization and is based in the United States. It covers a wide range of fields including computers, sustainable energy systems, aerospace, communications, robotics, and healthcare, and is composed of more than 420,000 members in more than 160 countries around the world.

About Couger Inc.

We are developing a human-like AI-based “Virtual Human Agent” that combines AI, IoT, AR/VR, and blockchain.

Web: https://couger.co.jp/
Contact us: contact@couger.co.jp