Computers will soon see and interpret images like humans - Tampere University of Technology

Computers will soon see and interpret images like humans

In January, Jiri Matas visited Tampere and the FiDiPro project was launched. The project aims at top-quality research, the development of publication activities and research results that are immediately applicable for businesses.

Expert in computer vision and machine learning, Professor Jiri Matas began his four-year FidiPro professorship at TUT’s Department of Signal Processing in January. His project focuses on the development of mobile technology that enables better and more diversified utilization of vast image data masses.

Jiri Matas comes from the renowned Center for Machine Perception laboratory at the Czech Technical University in Prague. In Tampere, he works in the Computer Vision Group led by Professor Joni Kämäräinen.

“My latest research areas have included image search, free text recognition, tracking and, most recently, deep neural networks. Another long-term interest of mine is real-time pattern recognition,” Matas lists.

For image search, Matas and his research team have sought new types of search criteria. With the methods applied today, the results produced by the searches include numerous nearly identical images of the same object.

“When the engine is also able to recognize images taken of the same object in different circumstances and at various distances, the resulting image selection can be more varied and a lot more informative.”

An engine that recognizes free text is able to scan a cityscape and read such items as advertising signs and licence plates. Tracking, on the other hand, refers to monitoring the movement of an object from one camera to another. It can be applied for surveillance purposes or as a part of a system for autonomous vehicles. A major breakthrough was recently achieved in the research on deep neural networks as the machine learned to identify and name objects in an image.

High-quality science yields useful applications

The partners of the Tekes-funded project include Finnish companies operating in telecommunications, software development and electronics development. The research targets, founded on practical problems, have been agreed upon together with the project partners.

Instead of applications per se, Matas highlights that the goal is to conduct high-quality research.

“All this may not sound very concrete and practical, but research in this field truly generates results that are applicable straight away.”

His research and business life cooperation so far have, in fact, generated a wide range of applications. For example, image search can be harnessed for tracing stolen goods placed on sale online. Facial recognition and analysis are particularly interesting for advertisers, for example. In time, free text recognition will enable applications that transcend the different alphabet systems and languages. They will benefit western travellers, for example, when in countries using some other alphabet system besides the Latin one.

New instructional methods and improved publications

Matas is renowned for his notable and visible results both as a researcher and an instructor. In Tampere, two doctoral students study under his guidance and he is also a mentor for the senior researchers in the group.

“We are intrigued to see what we can learn from his instructional methods. His example has also inspired us to hone our publication activities. Much of the focus in Finland seems to be on quantity instead of quality. We should retain our composure and invest in high-end research,” says Kämäräinen.

‘Pocket-Sized Big Visual Data’ is a consortium project with the Center for Machine Vision at the University of Oulu.

Updated by: Merja Jaaksi, 14.04.2016 14:49.
Content owner: Laukkanen Tuuli
Keywords: science and research, image search, mobile technology