人臉識(shí)別模塊的實(shí)現(xiàn)外文文獻(xiàn)翻譯、中英文翻譯、外文翻譯,識(shí)別,模塊,實(shí)現(xiàn),外文,文獻(xiàn),翻譯,中英文 譯文： 人臉識(shí)別模塊的實(shí)現(xiàn) 尾形克彥 傳統(tǒng)的人臉識(shí)別技術(shù)主要是基于可見(jiàn)光圖像的人臉識(shí)別，這也是人們熟悉的識(shí)別方式，已有30多年的研發(fā)歷史。但這種方式有著難以克服的缺陷，尤其在環(huán)境光照發(fā)生變化時(shí)，識(shí)別效果會(huì)急劇下降，無(wú)法滿足實(shí)際系統(tǒng)的需要。解決光照問(wèn)題的方案有三維圖像人臉識(shí)別，和熱成像人臉識(shí)別。但這兩種技術(shù)還遠(yuǎn)不成熟，識(shí)別效果不盡人意。近年來(lái)迅速發(fā)展起來(lái)的一種解決方案是基于主動(dòng)近紅外圖像的多光源人臉識(shí)別技術(shù)。它可以克服光線變化的影響，已經(jīng)取得了卓越的識(shí)別性能，在精度、穩(wěn)定性和速度方面的整體系統(tǒng)性能超過(guò)三維圖像人臉識(shí)別。這項(xiàng)技術(shù)在近兩三年發(fā)展迅速，使人臉識(shí)別技術(shù)逐漸走向?qū)嵱没? 人臉與人體的其它生物特征（指紋、虹膜等）一樣與生俱來(lái)，它的唯一性和不易被復(fù)制的良好特性為身份鑒別提供了必要的前提，與其它類型的生物識(shí)別比較人臉識(shí)別具有如下特點(diǎn)： 非強(qiáng)制性：用戶不需要專門配合人臉采集設(shè)備，幾乎可以在無(wú)意識(shí)的狀態(tài)下就可獲取人臉圖像，這樣的取樣方式?jīng)]有“強(qiáng)制性”； 非接觸性：用戶不需要和設(shè)備直接接觸就能獲取人臉圖像； 并發(fā)性：在實(shí)際應(yīng)用場(chǎng)景下可以進(jìn)行多個(gè)人臉的分揀、判斷及識(shí)別； 除此之外，還符合視覺(jué)特性：“以貌識(shí)人”的特性，以及操作簡(jiǎn)單、結(jié)果直觀、隱蔽性好等特點(diǎn)。 人臉識(shí)別系統(tǒng)主要包括四個(gè)組成部分，分別為：人臉圖像采集及檢測(cè)、人臉圖像預(yù)處理、人臉圖像特征提取以及匹配與識(shí)別。 1.人臉圖像采集及檢測(cè) 人臉圖像采集：不同的人臉圖像都能通過(guò)攝像鏡頭采集下來(lái)，比如靜態(tài)圖像、動(dòng)態(tài)圖像、不同的位置、不同表情等方面都可以得到很好的采集。當(dāng)用戶在采集設(shè)備的拍攝范圍內(nèi)時(shí)，采集設(shè)備會(huì)自動(dòng)搜索并拍攝用戶的人臉圖像。 人臉檢測(cè)：人臉檢測(cè)在實(shí)際中主要用于人臉識(shí)別的預(yù)處理，即在圖像中準(zhǔn)確標(biāo)定出人臉的位置和大小。人臉圖像中包含的模式特征十分豐富，如直方圖特征、顏色特征、模板特征、結(jié)構(gòu)特征及Haar特征等。人臉檢測(cè)就是把這其中有用的信息挑出來(lái)，并利用這些特征實(shí)現(xiàn)人臉檢測(cè)。 主流的人臉檢測(cè)方法基于以上特征采用Adaboost學(xué)習(xí)算法，Adaboost算法是一種用來(lái)分類的方法，它把一些比較弱的分類方法合在一起，組合出新的很強(qiáng)的分類方法。 人臉檢測(cè)過(guò)程中使用Adaboost算法挑選出一些最能代表人臉的矩形特征(弱分類器)，按照加權(quán)投票的方式將弱分類器構(gòu)造為一個(gè)強(qiáng)分類器，再將訓(xùn)練得到的若干強(qiáng)分類器串聯(lián)組成一個(gè)級(jí)聯(lián)結(jié)構(gòu)的層疊分類器，有效地提高分類器的檢測(cè)速度。 2.人臉圖像預(yù)處理 人臉圖像預(yù)處理：對(duì)于人臉的圖像預(yù)處理是基于人臉檢測(cè)結(jié)果，對(duì)圖像進(jìn)行處理并最終服務(wù)于特征提取的過(guò)程。系統(tǒng)獲取的原始圖像由于受到各種條件的限制和隨機(jī)干擾，往往不能直接使用，必須在圖像處理的早期階段對(duì)它進(jìn)行灰度校正、噪聲過(guò)濾等圖像預(yù)處理。對(duì)于人臉圖像而言，其預(yù)處理過(guò)程主要包括人臉圖像的光線補(bǔ)償、灰度變換、直方圖均衡化、歸一化、幾何校正、濾波以及銳化等。 3.人臉圖像特征提取 人臉圖像特征提?。喝四樧R(shí)別系統(tǒng)可使用的特征通常分為視覺(jué)特征、像素統(tǒng)計(jì)特征、人臉圖像變換系數(shù)特征、人臉圖像代數(shù)特征等。人臉特征提取就是針對(duì)人臉的某些特征進(jìn)行的。人臉特征提取，也稱人臉表征，它是對(duì)人臉進(jìn)行特征建模的過(guò)程。人臉特征提取的方法歸納起來(lái)分為兩大類：一種是基于知識(shí)的表征方法；另外一種是基于代數(shù)特征或統(tǒng)計(jì)學(xué)習(xí)的表征方法。 基于知識(shí)的表征方法主要是根據(jù)人臉器官的形狀描述以及他們之間的距離特性來(lái)獲得有助于人臉?lè)诸惖奶卣鲾?shù)據(jù)，其特征分量通常包括特征點(diǎn)間的歐氏距離、曲率和角度等。人臉由眼睛、鼻子、嘴、下巴等局部構(gòu)成，對(duì)這些局部和它們之間結(jié)構(gòu)關(guān)系的幾何描述，可作為識(shí)別人臉的重要特征，這些特征被稱為幾何特征?；谥R(shí)的人臉表征主要包括基于幾何特征的方法和模板匹配法。 4.人臉圖像匹配與識(shí)別 人臉圖像匹配與識(shí)別：提取的人臉圖像的特征數(shù)據(jù)與數(shù)據(jù)庫(kù)中存儲(chǔ)的特征模板進(jìn)行搜索匹配，通過(guò)設(shè)定一個(gè)閾值，當(dāng)相似度超過(guò)這一閾值，則把匹配得到的結(jié)果輸出。人臉識(shí)別就是將待識(shí)別的人臉特征與已得到的人臉特征模板進(jìn)行比較，根據(jù)相似程度對(duì)人臉的身份信息進(jìn)行判斷。這一過(guò)程又分為兩類：一類是確認(rèn)，是一對(duì)一進(jìn)行圖像比較的過(guò)程，另一類是辨認(rèn)，是一對(duì)多進(jìn)行圖像匹配對(duì)比的過(guò)程。 國(guó)籍：美國(guó) 出處：《人臉識(shí)別技術(shù)》普倫蒂斯霍爾出版社 原文： Face recognition technology Katsuhiko Ogata Traditional face recognition technology is mainly based on visible image face recognition, which is also a familiar way of recognition, has more than 30 years of research and development history. However, this method has some defects that are difficult to overcome. Especially when the ambient illumination changes, the recognition effect will drop sharply, which cannot meet the needs of the actual system. Solutions to solve the lighting problem include 3D image face recognition, and thermal image face recognition. However, these two technologies are still far from mature and the recognition effect is not satisfactory. In recent years, a rapidly developed solution is multi-light source face recognition technology based on active near-infrared image. It can overcome the impact of light changes, has achieved excellent recognition performance, in accuracy, stability and speed of the overall system performance than 3D image face recognition. This technology has developed rapidly in the past two or three years, making face recognition technology increasingly practical. Face and other biological characteristics of the human body (fingerprint, iris, etc.) as innate, its uniqueness and not easy to be copied good characteristics for identity identification provides the necessary premise, compared with other types of biometric face recognition has the following characteristics: Non-mandatory: the user does not need to specially cooperate with the face acquisition equipment, can almost in the unconscious state can obtain the face image, such a sampling method is not "mandatory"; Non-contact: users can obtain face images without direct contact with the device; Concurrency: in the actual application scenarios can be multiple face sorting, judgment and recognition; In addition, it also conforms to the visual characteristics: the characteristics of "knowing people by appearance", and the characteristics of simple operation, intuitive results, good concealment and so on. Face recognition system mainly includes four parts, respectively: face image acquisition and detection, face image preprocessing, face image feature extraction and matching and recognition. 1. Face image acquisition and detection Face image acquisition: different face images can be collected through the camera lens, such as static images, dynamic images, different positions, different expressions and other aspects can be very good collection. When the user is within the shooting range of the acquisition device, the acquisition device will automatically search and take the user's face image. Face detection: Face detection in practice is mainly used for face recognition pretreatment, that is, in the image to accurately calibrate the position and size of the face. Face image contains a very rich pattern features, such as histogram features, color features, template features, structural features and Haar features. Face detection is to pick out the useful information, and use these features to achieve face detection. Mainstream face detection methods adopt Adaboost learning algorithm based on the above features. Adaboost algorithm is a method used for classification. It combines some relatively weak classification methods together to create a new strong classification method. Face detection in the process of using Adaboost algorithm can pick out some of the most representative characteristics of face of the rectangular (weak classifier), according to the weighted voting to weak classifier structure for a strong classifier, and then obtained a number of strong classifier training series consists of a cascade structure cascade classifier, effectively improve the detection speed of classifier. 2 face image preprocessing Face image preprocessing: the face image preprocessing is based on the face detection results, the image processing and ultimately serve the process of feature extraction. The original image obtained by the system can not be used directly because of the limitation of various conditions and random interference. It must be preprocessed in the early stage of image processing, such as gray correction and noise filtering. For face image, its preprocessing process mainly includes face image light compensation, gray transformation, histogram equalization, normalization, geometric correction, filtering and sharpening. 3. Face image feature extraction Face image feature extraction: face recognition system can be used features are usually divided into visual features, pixel statistical features, face image transformation coefficient features, face image algebraic features, etc.. Face feature extraction is aimed at some features of the face. Face feature extraction, also known as face representation, is the process of face feature modeling. Facial feature extraction methods can be summarized into two categories: one is knowledge-based characterization method; The other method is based on algebraic features or statistical learning. The knowledge-based representation method is mainly based on the shape description of the face organs and the distance between them to obtain the feature data helpful to face classification, the feature components usually include the Euclidean distance between the feature points, curvature and Angle, etc. The face is composed of eyes, nose, mouth, chin and other parts. The geometric description of these parts and the structural relationship between them can be used as an important feature to recognize the face. These features are called geometric features. Knowledge-based face representation mainly includes geometric feature-based method and template matching method. 4. Face image matching and recognition Face image matching and recognition: the extraction of face image feature data and the database stored in the feature template search matching, by setting a threshold, when the similarity exceeds this threshold, the matching results output. Face recognition is to recognize the face features and the face features template has been compared, according to the degree of similarity on the face of the identity information to judge. This process is divided into two categories: one is confirmation, is a one-to-one image comparison process, the other is recognition, is a one-to-many image matching comparison process. Nationality: USA Source: Face recognition technology 譯文： 關(guān)于實(shí)現(xiàn)人臉?biāo)惴ㄗR(shí)別 人臉識(shí)別系統(tǒng)包括圖像攝取、人臉定位、圖像預(yù)處理、以及人臉識(shí)別（身份確認(rèn)或者身份查找）。系統(tǒng)輸入一般是一張或者一系列含有未確定身份的人臉圖像，以及人臉數(shù)據(jù)庫(kù)中的若干已知身份的人臉圖象或者相應(yīng)的編碼，而其輸出則是一系列相似度得分，表明待識(shí)別的人臉的身份。 人臉識(shí)別算法分類 基于人臉特征點(diǎn)的識(shí)別算法（Feature-based recognition algorithms）。 基于整幅人臉圖像的識(shí)別算法（Appearance-based recognition algorithms）。 基于模板的識(shí)別算法（Template-based recognition algorithms）。 利用神經(jīng)網(wǎng)絡(luò)進(jìn)行識(shí)別的算法（Recognition algorithms using neural network）。 神經(jīng)網(wǎng)絡(luò)識(shí)別 基于光照估計(jì)模型理論，提出了基于Gamma灰度矯正的光照預(yù)處理方法,并且在光照估計(jì)模型的基礎(chǔ)上，進(jìn)行相應(yīng)的光照補(bǔ)償和光照平衡策略。 優(yōu)化的形變統(tǒng)計(jì)校正理論基于統(tǒng)計(jì)形變的校正理論，優(yōu)化人臉姿態(tài)；強(qiáng)化迭代理論。強(qiáng)化迭代理論是對(duì)DLFA人臉檢測(cè)算法的有效擴(kuò)展； 獨(dú)創(chuàng)的實(shí)時(shí)特征識(shí)別理論 該理論側(cè)重于人臉實(shí)時(shí)數(shù)據(jù)的中間值處理，從而可以在識(shí)別速率和識(shí)別效能之間，達(dá)到最佳的匹配效果。 人臉識(shí)別的優(yōu)勢(shì)在于其自然性和不被被測(cè)個(gè)體察覺(jué)的特點(diǎn)。 虹膜識(shí)別 所謂自然性，是指該識(shí)別方式同人類（甚至其他生物）進(jìn)行個(gè)體識(shí)別時(shí)所利用的生物特征相同。例如人臉識(shí)別，人類也是通過(guò)觀察比較人臉區(qū)分和確認(rèn)身份的，另外具有自然性的識(shí)別還有語(yǔ)音識(shí)別、體形識(shí)別等，而指紋識(shí)別、虹膜識(shí)別等都不具有自然性，因?yàn)槿祟惢蛘咂渌锊⒉煌ㄟ^(guò)此類生物特征區(qū)別個(gè)體。 不被察覺(jué)的特點(diǎn)對(duì)于一種識(shí)別方法也很重要，這會(huì)使該識(shí)別方法不令人反感，并且因?yàn)椴蝗菀滓鹑说淖⒁舛蝗菀妆黄垓_。人臉識(shí)別具有這方面的特點(diǎn)，它完全利用可見(jiàn)光獲取人臉圖像信息，而不同于指紋識(shí)別或者虹膜識(shí)別，需要利用電子壓力傳感器采集指紋，或者利用紅外線采集虹膜圖像，這些特殊的采集方式很容易被人察覺(jué)，從而更有可能被偽裝欺騙。 人臉識(shí)別被認(rèn)為是生物特征識(shí)別領(lǐng)域甚至人工智能領(lǐng)域最困難的研究課題之一。人臉識(shí)別的困難主要是人臉作為生物特征的特點(diǎn)所帶來(lái)的。 相似性 人臉類似性 不同個(gè)體之間的區(qū)別不大，所有的人臉的結(jié)構(gòu)都相似，甚至人臉器官的結(jié)構(gòu)外形都很相似。這樣的特點(diǎn)對(duì)于利用人臉進(jìn)行定位是有利的，但是對(duì)于利用人臉區(qū)分人類個(gè)體是不利的。 易變性 人臉的外形很不穩(wěn)定，人可以通過(guò)臉部的變化產(chǎn)生很多表情，而在不同觀察角度，人臉的視覺(jué)圖像也相差很大，另外，人臉識(shí)別還受光照條件（例如白天和夜晚，室內(nèi)和室外等）、人臉的很多遮蓋物（例如口罩、墨鏡、頭發(fā)、胡須等）、年齡等多方面因素的影響。 在人臉識(shí)別中，第一類的變化是應(yīng)該放大而作為區(qū)分個(gè)體的標(biāo)準(zhǔn)的，而第二類的變化應(yīng)該消除，因?yàn)樗鼈兛梢源硗粋€(gè)個(gè)體。通常稱第一類變化為類間變化（inter-class difference），而稱第二類變化為類內(nèi)變化（intra-class difference）。對(duì)于人臉，類內(nèi)變化往往大于類間變化，從而使在受類內(nèi)變化干擾的情況下利用類間變化區(qū)分個(gè)體變得異常困難。 國(guó)籍：美國(guó) 出處：《關(guān)于人臉識(shí)別算法的實(shí)現(xiàn)》普倫蒂斯霍爾出版社 原文： On the realization of face algorithm recognition Katsuhiko Ogata Face recognition system includes image ingestion, face location, image preprocessing, and face recognition (identity confirmation or identity search). The system input is generally a piece or a series of face images containing unidentified identity, as well as a number of known identity face images in the face database or the corresponding coding, and its output is a series of similarity scores, indicating the identity of the face to be recognized. Face recognition algorithm classification Feature-based recognition algorithms based on face Feature points. Appearance-based recognition algorithms for whole face images. Template-Based Recognition Algorithms. Recognition algorithms using neural network. Neural network recognition Based on the theory of light estimation model, a light preprocessing method based on Gamma gray correction is proposed, and the corresponding light compensation and light balance strategies are carried out on the basis of the light estimation model. Based on the correction theory of statistical deformation, the face pose is optimized. Reinforce iteration theory. Enhanced iteration theory is an effective extension of DLFA face detection algorithm. Original real - time feature recognition theory The theory focuses on the intermediate value processing of real-time face data, so that the best matching effect can be achieved between the recognition rate and the recognition efficiency. The advantage of face recognition lies in its nature and the characteristics of not being detected by the individual. Iris recognition The so-called naturalness means that the identification method is the same as the biological characteristics used by human beings (or even other organisms) for individual identification. For example, face recognition, human is also through the comparison of face discrimination and identification, in addition to natural recognition and voice recognition, shape recognition, and fingerprint recognition, iris recognition, etc., are not natural, because humans or other organisms do not distinguish individuals through such biological characteristics. The quality of being unobservable is also important for a method of recognition, which makes the method not objectionable and, because it is not easy to draw attention to, less susceptible to deception. Face recognition has the characteristics of this aspect, it fully using visible light for face image information, and is different from the fingerprint recognition or iris recognition, need to use electronic pressure sensor fingerprinted, or by using infrared acquisition iris image, the mode of these special collection is very easy to detect, thus is more likely to be deceived by camouflage. Face recognition is considered to be one of the most difficult research topics in the field of biometric recognition and even artificial intelligence. The difficulty of face recognition is mainly caused by the characteristics of face as biological features. similarity Facial similarity There is little difference between individuals. All faces are structurally similar, and even the facial organs are structurally similar. This feature is beneficial to the use of human face localization, but it is unfavorable to the use of human face to distinguish human individuals. variability Facial appearance is not very stable, people can through the change of face a lot of expressions, and in different viewing Angle, the face of visual images also vary widely, in addition, face recognition is affected by light conditions (such as day and night, indoor and outdoor, etc.), face a lot of cover (for example, masks and sunglasses, hair, beard, etc.), the influence of various factors such as age. In face recognition, the first type of changes should be magnified as a standard to distinguish individuals, while the second type of changes should be eliminated because they can represent the same individual. Usually called the first type of change for inter-class change (inter-class difference), and called the second type of change for intra-class change (intra-class difference). For faces, the intra-class variation is often greater than the inter-class variation, which makes it extremely difficult to distinguish individuals by inter-class variation under the interference of intra-class variation. Nationality: USA Source: On the realization of face algorithm recognition