An Active Human-Machine Interface based on Multi-Depth of Field Face Recognition

• Main Authors: Chun-Chi Chang, 張鈞淇
• Other Authors: Jiann-Jone Chen
• Format: Others
• Language: zh-TW
• Published: 2012
• Online Access: http://ndltd.ncl.edu.tw/handle/39253260916107046579

碩士 === 國立臺灣科技大學 === 電機工程系 === 100 === An active face recognition system (AFR) is proposed to act as the real-time human-machine interface for an interactive TV (iTV) system. This AFR system is developed to recognize multi-persons with multi-depth of field for the iTV to recommend interested TV programs for current group users. It is assumed to operate in the family environment with four to six members and designed to recognize registered users and reject (identify) unregistered ones for the program recommender. To improve the 2D face image based recognition method, the 3D scene depth information is also acquired to eliminate false positive recognition. The Microsoft Kinect Sensor is utilized to provide the scene depth and body skeleton information to enhance the processing time and stability of the AFR. With the help of body skeleton, the face image area can be quickly located for the AFR to justify faces within which efficiently. The detected face is first filtered by an average face alpha mask for noise free and its scene depth is used to select corresponding face models from databases for recognition. For robust face feature description, we proposed to adopt multi-scale block local binary pattern, MBLBP. One face image is hierarchically decomposed into 1×1, 2×2, and 3×3 sub-blocks, 14 sub-blocks in total, from which the MBLBP features are extracted and whose histograms are concatenated to act as the face feature. The chi-squared between two sub-block histograms is calculated as the distance for dissimilarity measurement. Votes for these 14 sub-blocks matching between the unknown user and that in the database are counted for user recognition. In addition to face recognition, the AFR can also perform self-learning function to update face models. To achieve active face recognition, it has to recognize users under different lightings and different locations. Once recognized one user, the AFR select sub-blocks with medium similarity measure to update the corresponding ones in the face model. Experiments verified that the proposed AFR can achieve high accuracy and robustness in recognize multi-users with multi-depth of field.