【正文】 (412)二、對(duì)圖像進(jìn)行信息壓縮由于PCA使用了KL變換處理一個(gè)圖像序列，不僅去除了圖像信息間的相關(guān)性，同時(shí)降低了數(shù)據(jù)維度，保留了圖像信息的主成分。 將所有圖像排成一個(gè)N2*200的矩陣：ImageMatrix = (ImageVec1, ImageVec2, … , ImageVec200) (413)然后對(duì)它們進(jìn)行PCA處理，找出主元。測(cè)試圖像與訓(xùn)練圖像一般投影在變換后的主特征向量上，對(duì)所得投影值計(jì)算距離的接近程度劃分人臉類。PCA負(fù)責(zé)處理多元數(shù)據(jù)，它可以表示最佳的原始數(shù)據(jù)及其初始化，可以滿足最優(yōu)的最小均方誤差意義。第二節(jié) 讀入人臉庫(kù)建立人臉空間一、人臉庫(kù)簡(jiǎn)介實(shí)驗(yàn)中使用的是著名的ORL人臉庫(kù)[15]。 ORL人臉庫(kù)中前十幅人臉圖片 校正后的人臉庫(kù)中前十幅人臉圖片實(shí)驗(yàn)所用的ORL人臉庫(kù)每張46*56的圖片可以看成是一個(gè)2576維的向量，對(duì)ORL人臉庫(kù)，把一幅圖像按列優(yōu)先排成一維行向量（1*2576），從幾何空間角度來說，一個(gè)2576維空間中一點(diǎn)代表了一張人臉圖像。原始人臉圖像由張成新空間的正交基線性組合而成。計(jì)算出CA的非零特征值（降序排列，1=rM）及其對(duì)應(yīng)的特征向量后，可以得到新的特征子空間：投影后，可以得到計(jì)算訓(xùn)練集或測(cè)試集中人臉圖像在特征空間U中的坐標(biāo)： (52)第三節(jié) 特征向量的選取盡管求得的協(xié)方差矩陣已經(jīng)在很大程度上降低了維度，但是通常情況下，矩陣維度仍然很大，對(duì)于處理速度和識(shí)別效果兩方面的要求來說，大部分新的子空間坐標(biāo)成分仍對(duì)識(shí)別貢獻(xiàn)很??；另外，創(chuàng)建子空間所用的特征向量的數(shù)目與特征空間投影的計(jì)算速度有直接關(guān)系，在識(shí)別過程中沒必要保留所有的特征向量。這樣既能提高運(yùn)算速度，也不會(huì)影響識(shí)別效果。圖像被投影到子空間后，通常有兩種方法對(duì)相似性進(jìn)行判別：可以計(jì)算在投影子空間中圖像間的距離。本文選用的是三階近鄰分類算法。訓(xùn)練之后構(gòu)成200*200的特征矩陣，我們選取占90%“能量”的主成分，這個(gè)由主成分構(gòu)成的圖像子空間，任何一幅人臉圖像都可以向其投影得到一組投影值，這組投影值代表了該圖像在子空間中的位置，由距離度量標(biāo)準(zhǔn)我們可以確定圖片的接近程度。隨著維數(shù)的增多會(huì)導(dǎo)致最近鄰的期望距離急劇上升。第六節(jié) 本章小結(jié) 閱讀文獻(xiàn)并結(jié)合實(shí)驗(yàn)分析，我們可知，基于正弦變化的姿態(tài)校正算法對(duì)多姿態(tài)人臉識(shí)別是有效果的，它是一種簡(jiǎn)單，快速，實(shí)用的姿態(tài)校正算法。本論文所使用的PCA特征提取方法是一種簡(jiǎn)單有效的基于變換系數(shù)特征的算法，它的優(yōu)點(diǎn)主要有對(duì)質(zhì)量較高灰度人臉圖像可以直接拿來做訓(xùn)練集，識(shí)別時(shí)也可直接對(duì)圖像進(jìn)行投影分類。結(jié) 論本文對(duì)多姿態(tài)人臉識(shí)別算法中的姿態(tài)校正作了相關(guān)研究，對(duì)正弦變換的原理和效果進(jìn)行了一系列探索，并基于PCA的特征提取方法，實(shí)現(xiàn)了完整的識(shí)別過程。這種方法具有計(jì)算代價(jià)小、易實(shí)現(xiàn)等優(yōu)點(diǎn)，可以實(shí)現(xiàn)快速的人臉姿態(tài)校正，并且完整的保留了人臉圖像的紋理細(xì)節(jié)信息。如果圖像因?yàn)楸砬樽兓蛘哒趽跷镉绊戄^大情況下，人臉識(shí)別率不可能單純依靠某一算法得到保障。嘗試使用其他的特征提取方法和不同的分類器來做仿真實(shí)驗(yàn)。致 謝在此次畢業(yè)設(shè)計(jì)的完成過程中，我學(xué)到了很多東西。除此之外，與我同組研究人臉識(shí)別的谷春喜和張恩澤同學(xué)也給了我很大的幫助，由于初期對(duì)了我所要研究的課題不是特別了解，對(duì)此課題的開發(fā)語言也不是特別熟悉。 Multiview face detection using ’Haar’ Cascades, Face recognition using weighted modular PCA and Multiple Face Kalman Tracker.2 Related WorkTurk et. al came up with the statistical approach of describing faces in terms of the variation occurring among the faces of different individuals in the dataset [2]. In other words, selected number of Eigenvectors or Eigenfaces are puted from the set of images of different individuals using principal ponent analysis where these Eigenfaces span the feature space which maximizes the variation between the training faces. Recognizing of a face requires only a projection of this test face onto to this reduced dimensionality Eigenspace and pares the weights obtained to those of the faces trained. Pentland et al. extended their approach of Eigenfaces to include pose variations in the database and they proposed two methods [3]. One is to get a parametric Eigenspace which will encode both the face and pose variation. The other is a set of Eigenspaces with each one representating the variation of a subset of faceswith the same view. The appropriate Eigenspace for a test face can be determined by the distancefromfacespace metric[2] using each set of Eigenvectors. In addition to the Eigenfaces, other facial features such as eyes, noses and mouth were also coded to get Eigeneyes, Eigennoses and Eigenmouths. Their detection in a facial region is done by distancefromfeaturespace metric.An algorithm based on LDA was proposed by Etemad and Chellapa the problem was approached in a different than the one using PCA [4]. It focused on maximizing the separation of various face classes and minimizing the variance of faces images within a class rather than on finding a pact representation of face images like in PCA. Here, from the training images, the within class matrix and between class matrix are puted and bined to form the separation matrix. The face space or feature space is obtained by performing the Eigen analysis on this separation matrix. Performing PCA on the training set of images gives a basis set which separates pairwise relationships between pixels, more specifically the first and second order statistics. The higher order statistical relationship between pixels, which is the phase spectrum of the face image, is not captured by the PCA . Bartlett et al illustrated that the phase spectrum captures the structural information of the image that is more useful for face recognition rather than the amplitude spectrum captured by PCA [5]. Independent Component Analysis, which is a generalized version of PCA, attempts to capture not only second order statistics but higher order statistics corresponding to the phase spectrum and thus creates a set of basis images which are independent of each other.Liau et al. proposed an algorithm which is based on the ”viewbased” Eigen spaces where the view is incorporated by estimating the pose of the face in YCrCb colorspace [6]. Depending on the pose, the faces in the database are grouped and the mean face per group is puted. Using a similarity measure using the Euclidean distance, the test face is pared with each of the mean faces for pose estimation and then, extracts the suitable features using PCA. Only a global set of eigenvectors is used and the Eigenspace spans the variations of the face due to both the pose and facial features. However, it remains unclear how the pose estimation is used to discriminate the variations due to pose and the individual. The algorithm proposed in this paper uses a similar notion of ”viewbased” Eigenspaces mentioned in [3]. The database are grouped according to the pose as in [6] but each group has its own Eigenspace. Face recognition is done by projecting a detected face on to this selected Eigenspace and finding the closest match. In the next section, a theoretical overview and the implementation of the various system modules is explained.3 Frontal Face Recognition and extension to MultiPoseThe previous section explained briefly how a face is detected in an image irrespective of the identity of the individual. Once the faces are detected, the next step is to recognize the individuals whose faces are being detected and this requires the system to extract certain features from the face which will discriminate between individuals. For the face recognition system, the weighted modular Principal Component Analysis technique has been implemented[10]. In the first stage, suitable preproces