灰階人臉辨識之研究

Abstract

本論文是以特徵臉辨識法為基礎，結合三種前級處理器，包括使用傅立業譜補償臉部旋轉情況、邊界運算子正規化臉部大小、同態濾波器消除光源不均的影響，來建立人像辨識系統。論文中以ORL資料庫作為驗證及測試的標準，將原始資料庫及擷取臉部範圍後的資料庫應用於特徵臉辨識系統，探討兩者的辨識成果及影響因素，並分析直方圖均衡化與同態濾波器兩種消除光源不均的處理方法之效果及對特徵臉辨識法的辨識率之影響。經過實驗分析，本論文獲得以下結論，使用傅立業譜補償臉部旋轉情況、邊界運算子正規化臉部大小及同態濾波器消除光源影響等三種前級處理，對於特徵臉的辨識率將有所提升，而僅使用同態濾波器的原始ORL資料庫對於整體的辨識率更能提升7%以上。最後，本論文利用特徵臉辨識系統的統計特性及粗調與微調的概念，建構一個二級辨識系統，辨識率可達95%，實驗結果顯示二級辨識系統儘管耗費較多的辨識時間，但確實可有效提升辨識率，因此未來可針對此方向進行更深入的探討與研究。

A face recognition algorithm based on a preprocessing units and Eigenfaces approach is proposed in the dissertation. The preprocessing units consist of the following aspects: First, we use Fourier spectrum to detect and compensate the rotation of face image. Second, we apply Sobel edge operator to normalize the face in the image. Third, we utilize homomorphic filter to eliminate the effect of illumination. In the experiments of the dissertation, the ORL database is used as a standard test bench to verify the performance of the proposed face recognition system. By using original ORL database and clipped ORL database we are able to verify proposed algorithm in conjunction with the recognition result. In the present research, both histogram equalization method and homomorphic filter method have been analyzed to compare the effect on the face recognition system. Preliminary results indicated that by using the homomorphic filter to eliminate the illumination effect surely increase the recognition rate of the face recognition system.Note that by applying homomorphic filter to the ORL database can increase 7% in recognition rate . Finally, a two-stage Eigenfaces-based recognition system according to the statistical characteristic of the experiment result and the concept of combining coarse turn and fine turn is established to achieve 95% recognition rate. However, the two-stage system required longer time to obtain higher recognition rate. A future research in building an efficient second stage face recognition is suggested to complete the overall face recognition algorithm.