摘 要
人臉識(shí)別是圖像處理、模式識(shí)別和人工智能研究的重點(diǎn)領(lǐng)域之一，其目的是利用計(jì)算機(jī)根據(jù)人臉的特征來(lái)鑒別人物的身份，在商業(yè)、安全、身份認(rèn)證、法律執(zhí)行方面具有廣泛的應(yīng)用?；诮y(tǒng)計(jì)特征的人臉識(shí)別是最受關(guān)注的人臉識(shí)別技術(shù)之一，它克服了其它人臉識(shí)別方法的種種缺點(diǎn)，利用完備的統(tǒng)計(jì)學(xué)知識(shí)，根據(jù)人臉的統(tǒng)計(jì)特征就可有效地進(jìn)行人臉識(shí)別。但是由于人臉模式的復(fù)雜性和多變性，在姿態(tài)、光照和表情等條件變化下人臉識(shí)別率會(huì)嚴(yán)重下降。因此，在特征提取中注重魯棒性，同時(shí)兼顧識(shí)別效率的人臉特征提取技術(shù)是當(dāng)前研究的熱點(diǎn)。
本文主要針對(duì)人臉特征提取技術(shù)進(jìn)行了研究，研究的重點(diǎn)是將全局特征方法與局部特征提取方法相結(jié)合，目的是在提高人臉識(shí)別率的基礎(chǔ)上，針對(duì)姿態(tài)、光照和表情等條件變化，進(jìn)一步提高人臉特征提取和識(shí)別算法的魯棒性。本文的主要工作和創(chuàng)新點(diǎn)如下：
(1) 在研究了局部二值模式算法的基礎(chǔ)上，將其與主成分分析算法相結(jié)合，提出了一種基于(2D)2PCA-LBP的人臉識(shí)別方法。該方法在提取了人臉紋理特征信息的基礎(chǔ)上，用(2D)2PCA方法進(jìn)行降維。LBP算法具有旋轉(zhuǎn)不變性，對(duì)光照變化和姿態(tài)變化具有一定的魯棒性；(2D)2PCA是PCA算法的改進(jìn)，可以對(duì)圖像進(jìn)行最大程度的降維。實(shí)驗(yàn)結(jié)果表明，該算法可以提高人臉識(shí)別率，并且對(duì)光照、姿態(tài)和表情變化有一定的魯棒性。
(2) 研究了壓縮傳感算法，針對(duì)人臉識(shí)別對(duì)遮擋、表情和光照等因素的魯棒性問(wèn)題，提出了一種基于PCA特征基壓縮傳感算法的人臉識(shí)別方法。該方法首先利用(2D)2PCA方法將人臉圖像變換到PCA特征域，將提取的圖像行列兩個(gè)方向的特征作為壓縮傳感算法的超完備基；然后通過(guò)求解最小化l1范數(shù)，尋求圖像在該超完備基上的稀疏表示，以得到一組最優(yōu)的稀疏系數(shù)來(lái)重構(gòu)各類圖像，通過(guò)求取測(cè)試圖像與重構(gòu)圖像的最小殘差進(jìn)行分類識(shí)別。該方法突破了傳統(tǒng)方法僅用一類訓(xùn)練樣本進(jìn)行識(shí)別的缺陷，將所有訓(xùn)練樣本同時(shí)用來(lái)進(jìn)行分類，分類效果得到改善，識(shí)別率顯著提高；同時(shí)，將時(shí)間復(fù)雜度降低到了線性階。研究表明，只要訓(xùn)練樣本足夠充分，就能有效地表示測(cè)試樣本的所有情況，在姿態(tài)、光照、表情變化比較大的情況下，識(shí)別率得到明顯改善。

關(guān)鍵詞 人臉識(shí)別；主成分分析法；局部二值模式；壓縮傳感；稀疏表示


Abstract
Face recognition is one of the most important research fields of image processing, pattern recognition and artificial intelligence, its purpose is to identify the identity of the people using computer through face feature. It has a wide range of applications, including commerce, security, person verification, and law enforcement. Face recognition based on statistical characteristics is now one of the most concerned face recognition technology by the researchers, it overcomes the shortcomings appeared in the process of traditional face recognition methods, using statistical knowledge can be effective for face recognition. However, since face patterns are complicated and multiform, the face recognition rate will face sharp decline under various conditions, such as changing illumination, pose and facial expression, which are very difficult to represent face effectively. Therefore, face recognition methods with robustness and efficiency are hotspots in recent studies.
The thesis studies on feature extraction problem in face recognition, and the focuses are the combination of global feature and local feature. Our goal is to further improve the robustness with varying expression, illumination and shadow while improving the face recognition rate at the same time. The main work and innovations are as follows:
1. This paper analyzes LBP algorithms, then combines it with PCA algorithm, presents a (2D)2PCA-LBP algorithm of face recognition. This method first extracts facial texture feature, then uses (2D)2PCA algorithm to reduce its dimension. The reason is that LBP algorithm has a characteristic of rotation invariance, which is robustness to illumination changing and pose variation. (2D)2PCA algorithm is the improvement of PCA, using this method, the image can reach the maximum degree of dimensionality reduction. The experiment results show that the algorithm can improve the face recognition rate, especially to the image with illumination changing, pose variation and facial expression, the recognition rate improve significantly.
2. The thesis studies the compressed sensing method in-depth. In order to solve the robustness problem with block, expression and illumination in face recognition system, we propose a face recognition method based on PCA-based compressed sensing algorithm. Utilizing (2D)2PCA transform to extract image features in both row and column directions and reducing the dimension. A projection matrix is constructed to identify the face features, considering these features to form an over complete dictionary. By solving the l1 norm minimization, seeking out the sparsest representation of images based on the dictionary to obtain a set of optimal sparse coefficients, which are used to recover the train images, compute the residuals between test and train images for face recognition. The method breaks through the characteristics of traditional method using only one class for recognition, we use all the training set for classification. The classifications results are improved and the time complexity is reduced to linear order, in the same time, the recognition rate improves..