E-mail r-ishiyama [at] nec.com, h-miyano [at] nec.com
To enable machines (Artificial Intelligence, AI) to work in in good harmony with humans, we are involved in research and education of technology for precise real-time recognition and comprehension using sensors such as cameras, especially of real-world situations where there are many people and objects moving around and interacting.
In recent years, technological innovations based on deep learning techniques have dramatically increased the performance of AI, particularly with regard to image recognition. It is expected that this technology will be used in diverse applications including real-time analysis of security camera footage, and inspection/robotics in factories. However, AI currently requires not only large amounts of learning data to be prepared in advance, but also large amounts of adjustments to adapt to each installation site. As a result, there are still many issues to be overcome in order to apply AI to diverse real environments that change from one moment to the next. Sensing more detailed data from the real world is the key technology.
To adapt to environmental changes, it is useful to capture changes in real-world conditions with faster real-time performance and in greater detail. In particular, if it is possible to perform not only spatial analysis of subjects that are targeted by most deep learning models, but also the detailed temporal analysis and comprehension, then it should be possible to grasp changes more reliably and adapt more easily to diverse environments. Specifically, at our laboratory we are mainly working on the following themes, but we also work on a wide variety of general recognition technologies primarily involving image recognition, such as improvements of deep learning itself.
High-speed-camera object recognition
Until recently, most image recognition studies have assumed a 30 fps frame rate (30 pictures captured per second). However, we aim to gain a deeper understanding of the real world by using a high-speed camera to obtain data with greater detail in the time axis (from 100 to 1000 fps) so that even fast-moving objects can be reliably tracked and evaluated without disturbing their motion, and so that tiny vibrations of objects can also be analyzed. This object recognition technology using high-speed cameras can achieve high speed inspection in, for example, the production of many models in small quantities where many different items are handled and each one has to be checked appropriately without interrupting the production process.
Individual object authentication
If it is possible to distinguish each individual item in a single camera image, then these items can be reliably tracked without having to perform constant sensing, and changes can be analyzed as they occur. With this as a broad theme, our aim is to individually identify and track any item in the real world by instantly capturing images with a camera and analyzing their detailed patterns instead of relying on special tags such as RFIDs. This will make it easy to find inefficiencies and optimize productivity, even in high-mix, low-volume production environments that are constantly changing, for example.
Joint research and collaboration
We are continuing to strengthen our core technological ability while promoting joint research with various research institutions including the University of Tokyo and the RIKEN Center for Advanced Intelligence Project (AIP).
Open and global research environmentt
We invite many researchers and internship students from Europe, Oceania and Asia to the open laboratories at NEC. Students of our laboratory learn about various research fields and languages, while gaining a global point of view.