Outline

To make it easy for people to search for scenes in which a particular actor appears, we are researching a technique that automatically adds semantic information (metadata) to TV program video. Here, we are exhibiting a technique that can automatically recognize the people appearing in a TV program.

Features

Face recognition that adapts robustly to TV programs

By using image features that take the rough positional relationships of facial landmarks such as the eyes and nose into consideration, we have made it possible to recognize faces with high precision even in a TV program where there are large variations in the lighting of people’s faces, their facial expressions, and the orientation of their faces.

An efficient method suited to the processing of large quantities of video data

By connecting a chain of decision processes and reducing the computational load by early rejecting regions that clearly do not contain faces before performing the detailed detection process, we have made it possible to process large quantities of video data such as NHK archives.