거리별 얼굴 영상을 이용한 실시간 원거리 얼굴 인식에 대한 연구

Abstract

The recent surveillance system is evolving into an intelligent system where it judges on its own and acts based upon image analysis technology. To satisfy the needs of high level intelligent surveillance system, it shall not only be able to extract objects and classify but also to identify precise information on the object. The representative method to identify one’s identity in the field of surveillance is face recognition. Since recently face recognition has been applied to wide distance of applications including criminal investigation, searching for a missing child, and Human-Robot Interaction(HRI), the needs for long distance face recognition system has been dramatically increased. Since the conventional face recognition system is applied to short distance face recognition such as access control, the need for long distance face recognition has not been considered. In other words, if conventional face recognition system is to be applied to surveillance, it may be hard to expect satisfying quality on long distance face recognition with low image resolution by changes in camera resolution and distance. Recently researches on long distance face recognition technology based on low resolution face information extracted from low resolution images or changes in distance haven been in progress. As representative methods of the foregoing, there are super resolution image reconstruction method, method using structural characteristics of images, face recognition using Pan Tilt Zoom(PTZ) camera. However, these conventional methods also have problems in which low recognition rate under changes in lighting and distance together with low processing speed for reconstruction of face images with low resolution, and dependability of learned data. In addition, the method using PTZ camera improves the rate of recognition, but does not solve fundamental issues in long distance face recognition such as additional costs. Thus, long distance face recognition technology, where it can be operable under environment with conventional low resolution and high resolution Closed Circuit TeleVision(CCTV), and possess strong real-time processing capability with changes in lighting and distance, is needed without additional costs. This paper propose a method that face images by actual distances shall be used in the stage of registering user’s face images for the purpose of real-time long distance face recognition. In order to remedy low recognition rate incurred by changes in distance, we would like to propose learned data composition method using face images by distances. This method using face images by distances can be classified into the followings depending upon the manner to acquire face images, for instance, face images by actual distances, face images by virtual distances using zoom camera, and face images by virtual distances using automatic creation method. The proposed method in comparison with the conventional method, which only acquires short distance face information, has merits in acquiring precise characteristics of individual faces in both short and long distance. The conventional method of long distance face recognition uses reconstruction method turning face images of long distance with low resolution into higher resolution whereas the proposed method is to make long distance face images with low resolution using short distance face images with high resolution through deterioration of image quality. In addition, the problems of nonconformity in face size and brightness of face by changes in distance can be solved by bilinear interpolation and histogram equalization respectively. The proposed method, that is learned data composition method using face images by distances, in this paper has 98.3% face recognition rate on average for entire distance in QVGA resolution achieving 58.7% higher than the conventional method using structural characteristics whereas 93.7% face recognition rate on average for entire distance in VGA resolution achieving 40.6% higher than the conventional method. Finally, the proposed method achieved 36.2% higher than the conventional method in SXGA resolution as well.