日夜視訊之人臉辨識

Abstract

生活中很多地方適合應用人臉辨識系統，諸如家庭照護與安全監控系統等等，其在全天候與即時系統中，常用來做影片的人臉辨識。為此，我們使用單台近紅外線攝影機，用於擷取日夜人臉影像進行人臉辨識。 此篇論文中，我們採用OpenCV做人臉偵測，其為使用Haar疊層分類器，是一種基於特徵運算的演算法，其較基於逐點的演算法更為快速。而在光線較昏暗的夜晚或無光之室內，攝影機拍攝之影像色調與對比度不如光線充足之白天，故本論文中，先將夜晚所拍攝到的影像以多尺度Retinex增強，再執行人臉擷取。 OpenCV的人臉擷取方法迄今為止非常受歡迎，但有其瑕疵。在擷取出來的人臉影像中，可能出現錯誤的非人臉影像。本篇論文中，我們將距離人臉群中心過遠的影像分類成非人臉影像。其可彌補OpenCV之缺點，並大幅提升人臉辨識率。 接著擷取出來的人臉影像經由特徵空間與標準空間轉換，累積三張上述人臉影像後，經由多數決的方式，完成人臉辨識。此外，我們加入兩位外來者之人臉影像，他們不在人臉影像之群中心訓練成員內，經測試此篇論文之人臉辨識架構，對外來者的辨識而予阻擋是有效的。

Human face recognition system is a desired technique in our daily life, such as the home nursing care system, security applications, and many others. It is a widely well-come technique that all-day-long and on-line to recognize a person from video camera. To this end, we use a near infrared (NIR) cameras to capture day-and-night video images for on-line human recognition. In this thesis, we adopt human face sub-image attraction package in OpenCV, which is based on Haar cascade classifier. The package is a feature-based algorithm and works much faster than the pixel-based algorithm. Moreover, the image contrast color tones of video frames in the night is worse than in the day, thus we employ multi-scale retinex to enhance video frames in the night before face extraction. Despite OpenCV’s popularity to date, the technique to extract face sub-images from taken videos are not reliable. We can obtain many non-face sub-images among those obtained extracted face sub-images. We judiciously collect extracted sub-images those very far-away from to the centroids of persons to be classified and then remove them as non-face sub-images. This may remedy the shortcoming of OpenCV package, and greatly increase the face recognition rates. Then the extracted face sub-image is transformed to a new space by eigenspace and canonical space transformation. The recognition is finally done in canonical space. Moreover, we consider most recent three consecutive face image recognition from video, and use majority vote to recognize the person. Moreover, we test face image recognition on two intruders, who do not belong to the members in the training set. Our proposed system can reject the intruder successfully.