熱人臉影像辨識之特徵抽取與分類器比較分析

A Comparative Analysis of Thermal Infrared Face Images Feature Extractions and Classifiers

莊宜叡(Yi-Jui Chuang)

指導教授 ： 劉益宏

中原大學/工學院/機械工程研究所/碩士(2009年)

https://doi.org/10.6840/cycu200901418

摘要

自動化人臉辨識、表情辨識系統等人身安全相關的新興科技在近代中受到愈來愈多的重視，也吸引了許多學者從事相關的研究。然而在可見光人臉辨識系統中常常會因為光源改變直接影響了分類的成效，所以可見光人臉辨識需要控制光源因素，才能有良好的成效。本文將過去文獻中所使用的特徵抽取方法(主成份分析Principal Component Analysis、線性鑑別分析Linear Discriminant Analysis、廣義鑑別分析Generalized Discriminant Analysis)且加入了核心主成份分析(Kernel Principal Component Analysis)運用到熱人臉影像做統合性的比較。在本文中，我們加入了支持向量機(SVM)分類器與先前所使用的分類器(最近平均分類器和K個最近鄰居分類器)進行比較。由特徵抽取比較實驗結果顯示，由分類成效可知資料點先經由核函數映射至高維特徵空間中進行特徵抽取，比直接在輸入空間中進行特徵抽取來得好。在分類器比較方面，將SVM運用在熱人臉影像辨識所得到的分類成效均比其他的分類器好，同時也證明直接運用熱影像的灰階值也可得到很高的分類率。本文中建立了中原大學熱紅外人臉影像資料庫，其中包含50個人，每個人60張熱影像照片，總共包含3000張熱影像照片。每人隨機挑選10張影像進行實驗。由實驗結果可得知對於包含人臉輪廓的資料分別運用主成份分析、核心主成份分析、線性鑑別分析和廣義鑑別分析分類率可高達100%，其中分類器為SVM。對於未包含人臉輪廓的資料使用SVM分類器，配合廣義鑑別分析分類率為99.2%，而配合核心主成份分析分類率可高達100%。

關鍵字

廣義鑑別分析 ； 線性鑑別分析 ； 核心主成份分析 ； 主成份分析 ； 支持向量機器 ； 熱人臉影像辨識

並列摘要

Fully automatic face and expression recognition systems have received increasingly attention in recently years. However, the classification performance in the visible light face recognition system is often directly affected by the light source changed. Therefore, the visible light face recognition must control the factors of the light source to get the better results. This thesis used the previous feature extraction methods (Principal Component Analysis, Linear Discriminant Analysis, and Generalized Discriminant Analysis) and joined the Kernel Principal Component Analysis to apply to the thermal face images to make the comprehensive comparison. In this thesis, we joined the Support Vector Machine (SVM) classifier and the previous classifiers (Nearest Mean Classifier and K Nearest Neighbor Classifier) for the comparison. Comparing the SVM classifier with the others ,the classification performance using it in the thermal imagery face recognition is better than the others, and the result also establishes that directly using the gray-level values of the thermal images to classify can get higher classification rates. This thesis establishes the thermal infrared face image database of Chung Yuan Christian University which includes 50 individuals, each person 60 thermal images, totally 3000 images. The results show that using the images containing face contours using PCA, KPCA, LDA and GDA respectively and using the SVM classifier can achieve 100% of the classification rate. For the images do not contain the face contours, the classified rate using the SVM classifier with GDA is 99.2 percent, and with KPCA can reach as higher as 100%.

並列關鍵字

kernel-based Generalized Discriminant Analysis ； Linear Discriminant Analysis ； Kernel Principal Component Analysis ； Principal Component Analysis ； IR Thermal Face Recognition

參考文獻

[41]陳衍廷，應用核心主成份分析及非平衡式SVM於影像中人臉偵測之研究， 私立中原大學機械工程研究所碩士學位論文， 2006.

[1]S. R. Hunter, G. Mauter, L. Jaing, and G. Simelgor, “High-sensitivity uncooled microcantilever infrared imaging arrays,” Proeeding SPIE, vol.6206, pp. 1-12, 2006.

[2]F. J. Prokoski, “History, current status, and future of infrared identification, ” in IEEE Workshop on Computer Vision Beyond the Visible Spectrum, Hilton Head, 2000.

[4]D. Socolinsky, A. Selinger and J. Neuheisel, “Face recognition with visible and thermal infrared imagery, ” Computer Vision and Image Understanding, vol. 91, pp. 72–114, 2003.

[5]D. Socolinsky and A. Selinger, “Thermal face recognition in an operational scenario, ” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2004.