基於階層式特徵法應用於不穩定光源下的人臉偵測系統

淡江大學機構典藏 > 工學院 > 電機工程學系暨研究所 > 學位論文 > Item 987654321/88104

請使用永久網址來引用或連結此文件: https://tkuir.lib.tku.edu.tw/dspace/handle/987654321/88104

題名:	基於階層式特徵法應用於不穩定光源下的人臉偵測系統
其他題名:	Hierarchical feature method for illumination variant face detection system
作者:	林進益;Lin, Chin-Yi
貢獻者:	淡江大學電機工程學系碩士在職專班余繁
關鍵詞:	人臉偵測;類神經網路;紋理特徵;即時偵測;Face detection;Adaboost;Neural Network;Texture feature;Real-time detection
日期:	2012
上傳時間:	2013-04-13 12:00:55 (UTC+8)
摘要:	一般boosting人臉偵測演算法使用矩形特徵。為了得到較好的效能，通常需要較多的訓練樣本，這會產生無法預估的特徵數量，可能需要更多的時間來偵測人臉。此外，矩形特徵是採用像素值來計算，像素值容易受到光源的影響，因此在計算特徵值前須做前處理的動作但也會增加其運算時間。本論文提出以Adaboost演算法和類神經網路法中的倒傳遞網路為基礎，搭配區域和全域特徵以串接式架構來偵測人臉。基於階層式的人臉偵測系統，具有即時偵測且對於光源低敏感度的特性。其中區域特徵我們採用MCT特徵(Modified Census Transform Feature)，是一種紋理特徵，對光源較不敏感。使用這類紋理特徵不用對每個子視窗作前處理的動作。在挑選弱分類器上採用與一般boosting演算法不同的方法，以階層式特徵架構來控制特徵數量以減少產生過多不必要的特徵。因為MCT特徵只有描述紋理資訊，亮度資訊被移除。若單以MCT特徵計算，容易造成誤判。所以在本論文中加入全域特徵，考量亮度資訊，可以排除很多誤判區域。實驗結果顯示本論文所提架構的偵測率為99%，誤判個數11個，偵測速度為27.92 FPS。 General boosting algorithms for face detection use rectangle features. To get better performance, it needs more training samples and may generate some unpredictable number of features and that is why it needs more time to detect the face. Besides, using pixel values, which are easily affected by illumination, to calculate the rectangle features, it usually needs to preprocess the data before calculating the values of features. Such approach may increase the computation time. Our proposed solution is based on Adaboost algorithm and back propagation network of neural network combining local and global features with cascade architecture to detect human faces. This thesis presents a hierarchical face detection system with real-time operation and low sensitivity to light illumination. We use Modified Census Transform Feature (MCT), which is belonged to texture features and is less sensitive to illumination, for local feature calculation. By this approach, it is not necessary to preprocess each sub-window of the image. For classification, the selection of weak classifiers is different from that of the boosting algorithms, here we use the structure of hierarchical feature to control the number of features and it will not generate too many of them. Since MCT only describes the texture information, the brightness information is removed. With only MCT, it is very easy to misjudge faces. Therefore, in this work we include the brightness information of global features to eliminate the false positive regions. The experimental results show that the proposed architecture can have detection rate of 99%, false positives of 11, and detection speed of 27.92 FPS.
顯示於類別:	[電機工程學系暨研究所] 學位論文

文件中的檔案:

檔案	大小	格式	瀏覽次數
index.html	0Kb	HTML	297	檢視/開啟

在機構典藏中所有的資料項目都受到原著作權保護.

TAIR相關文章

資料載入中.....