100 Years of Engineering HKU 100 Faculty of Engineering 100 Years of Engineering

HKU Engineering Centenary Exhibition

Objectives 研究目標
Performance of face recognition system will be severely degraded if input face is not frontal. To overcome this, we transformed a face with arbitrary pose to a frontal form that is suitable for face recognition.
當輸入的人臉圖象爲非正面時,人臉識別系統的性能將會急劇下降。爲了突破上述局限,我們將輸入圖象中任意姿態的人臉轉化爲適合人臉識別系統的正面姿態。

Overall Diagram 總體框圖

Face detection overall diagram

Methodology 主要算法

(1)

Facial Feature Detection
臉部特徵檢測

Uses AAM with minimum residual error to locate detailed facial features
利用具有最小殘差的主動外觀模型定位詳細的臉部特徵

(2)

3D Modeling & Rendering
三維建模和繪製

Employed 3D face mesh has 6292 vertices and 6152 facet, whose appearance could be modified by adjusting parameters
使用包含了6292個頂點及6152個面的三維網格,並且可以通過調整參數的方法來修改三維人臉的外觀

(3)

Pose estimation
姿態估計

Uses Moore-Penrose pseudo inverse to calculate the transformation matrix
利用Moore-Penrose 僞逆計算變換矩陣

(4)

Texture Map Rebuilding
紋理圖重建

Based on (3), texture in input image was filled and interpolated on texture map by texture mapping. A post processing is then performed for more realistic 3D face reconstruction results
根據步驟三之姿態估計結果,輸入圖象中的紋理被填充和插值到紋理圖中。然後使用一個後處理過程使得重建的人臉看起來更加真實。

Reconstruction Results 重建結果
Face images are tested with different poses in the GTAV database, non-frontal view face can be synthesized to a realistic frontal view
使用GTAV數據庫中不同姿態的人臉進行測試,非正面的人臉可以被合成爲正面人臉

Reconstruction Results

Contributions 研究成果
Digital video surveillance technologies are now becoming more and more mature and thus the entry barrier for this industry is becoming lower. We believe the technologies developed in this project, which integrates well with commodity-off-the-shelf face recognition package, can realize many value-added features for Hong Kong manufacturers to differentiate themselves from the other competitors around the globe.
數字視頻監控技術的日益成熟使得進入這項產業的門檻變低。我們相信此方案所開發的技術可與其他人臉識別系統相配,實現許多增值特性。並使得香港的相關產業區別於全球其他地區的競爭者。

Team Members 研究成員
Department of Computer Science: Dr CHOW Kam Pui, Prof CHIN Yuk Lun Francis, Dr WONG Kwan Yee Kenneth
計算機科學系: 鄒錦沛博士、 錢玉麟教授、黃君義博士

Project Sponsor 贊助機構
Innovation and Technology Fund, Innovation and Technology Commission 創新科技署創新及科技基金
CyberView Inc. Ltd.

This project is supported by Cognitec Systems Ltd. which uses its FaceVAC software to evaluate our proposed menthod
此項目由Cognitec Systems Ltd. 支持並使用其FaceVACS人臉識別軟件作測試

Video 短片
http://www.engineering.hku.hk/enggke/video.php?id=29