本文選題：人臉定位　切入點：人臉識別　出處：《貴州民族大學(xué)》2017年碩士論文　論文類型：學(xué)位論文

【摘要】：為了方便出入口人物身份的識別,減輕安保人員值班的壓力,前人在視頻監(jiān)控領(lǐng)域?qū)Τ鋈肟谌宋锷矸葑R別系統(tǒng)進行了大量探索,取得不少的成果,然而仍有一些問題值得進一步研究,例如:如何在人臉定位算法獲取定位結(jié)果的基礎(chǔ)上再次運用其它定位算法來確保定位的精度;怎樣在人臉識別精度與識別速度上做取舍;如何能隱性自動提取圖像特征。自適應(yīng)增強算法(Adaptive Boosting Algorithm,Adaboost算法)是常用的人臉定位算法,卷積神經(jīng)網(wǎng)絡(luò)(Convolutional Neural Network,CNN)是當(dāng)今深度學(xué)習(xí)的關(guān)鍵技術(shù)之一。本文在前人工作基礎(chǔ)上,采用Adaboost算法與CNN相結(jié)合開發(fā)出入口人物身份識別系統(tǒng),本系統(tǒng)主要功能是對出入口行人進行實時定位、特征提取、人物身份識別,實時掌握進入出入口人員身份,當(dāng)系統(tǒng)發(fā)現(xiàn)有外部人員出入時,本系統(tǒng)自動警示值班安保人員注意。本文的開發(fā)工作主要分為三個方面:1.使用高斯混合模型提取感興趣的運動目標(biāo)區(qū)域,為下一步的人臉定位有效地縮小了檢測范圍。2.提出一種結(jié)合Adaboost算法與CNN的人臉定位方法。該方法首先在目標(biāo)區(qū)域內(nèi)使用Adaboost算法尋找人臉區(qū)域;然后,把人臉圖像區(qū)域讀進CNN里再次判別,這樣可在Adaboost算法獲取定位結(jié)果的基礎(chǔ)上進一步提高人臉定位精度。3.提出一種采用步長為2的方法進行計算卷積層;減少網(wǎng)絡(luò)采樣層,在損失一定的識別精度的前提下,減少了人臉識別時間。將上述3點集成系統(tǒng)應(yīng)用于視頻監(jiān)控的出入口場景中,并對出入口人員進行身份識別。本文測試工作主要從人臉定位、人臉識別、系統(tǒng)實地場景三個方面進行測試:1.在美國加州理工學(xué)院吳恩達等人建立的人臉定位數(shù)據(jù)庫和實地場景進行人臉定位實驗表明,采用Adaboost算法與CNN相結(jié)合的方法在人臉定位精度上高于單純采用Adaboost的方法。2.在人臉識別公開庫yale與ORL測試實驗結(jié)果表明,改進的CNN算法雖然犧牲了一定的識別精度,但在識別時間上平均每張圖像識別時間減少了0.013秒。3.在貴州民族大學(xué)辦公樓、實驗樓、宿舍樓、幼兒園出入口實地場景進行測試。本系統(tǒng)在CNN訓(xùn)練過程時直接用原始圖像來訓(xùn)練,隱性地自動提取出圖像信息特征并分類,在人臉定位方面使用Adaboost算法和CNN相結(jié)合的方法有效剔除偽人臉圖像,為下一步的工作創(chuàng)造條件;而人臉識別方面采用改進的CNN算法,在犧牲一定的識別精度的前提下,減少了識別時間,為實時識別打下基礎(chǔ),該系統(tǒng)可以減輕安保人員的工作壓力和保障人們的人身及財產(chǎn)安全。因此,具有一定的應(yīng)用價值與推廣價值。
[Abstract]:In order to facilitate the identification of people at the entrance and exit, and to alleviate the pressure on security personnel on duty, the predecessors have made a great deal of exploration in the field of video surveillance and achieved a lot of results. However, there are still some problems worthy of further study, such as: how to use other localization algorithms to ensure the location accuracy again on the basis of the face localization algorithm to obtain the localization results, how to make a choice between the face recognition accuracy and the recognition speed, and how to make a choice between the face recognition accuracy and the recognition speed. Adaptive Boosting algorithm (Adaboost) is one of the most popular facial localization algorithms, and Convolutional Neural Network (CNN) is one of the key techniques of deep learning. The main function of this system is to locate the entrance pedestrian in real time, extract the feature, recognize the character identity, and master the identity of the person entering the entrance and exit in real time, by using Adaboost algorithm and CNN, and the main function of the system is to locate the pedestrian in real time. When the system detects the external personnel, the system automatically warns the security personnel on duty. The development work of this paper is divided into three aspects: 1.Using Gao Si mixed model to extract the moving target area of interest. For the next step, the detection range of face location is reduced effectively. 2. A face localization method combining Adaboost algorithm with CNN is proposed. Firstly, the Adaboost algorithm is used to find the face region in the target area. The face image region can be read into CNN again, so that the accuracy of face location can be further improved on the basis of Adaboost algorithm. 3. A method with step size of 2 to calculate convolutional layer is proposed to reduce the network sampling layer. On the premise of losing certain recognition accuracy, the time of face recognition is reduced. The above three-point integrated system is applied to the entrance scene of video surveillance, and the identification of the entrance and exit personnel is carried out. Face recognition and system field scene are tested in three aspects: 1.The experiments of human face location database and field scene established by Wu Enda and others of California Institute of Technology, USA show that, The combination of Adaboost algorithm and CNN method is more accurate than Adaboost method. The experimental results of yale and ORL in the face recognition open library show that the improved CNN algorithm sacrifices some recognition accuracy. However, the average recognition time per image was reduced by 0.013 seconds. 3. In the office building, experimental building, dormitory building of Guizhou University for nationalities, In the process of CNN training, the system directly uses the original image to train, and recessive automatically extracts the image information features and classifies them. In the aspect of face location, the method of combining Adaboost algorithm with CNN algorithm is used to eliminate the pseudo-face image effectively, which creates the conditions for the next step, while the improved CNN algorithm is used in face recognition, at the premise of sacrificing certain recognition accuracy. It can reduce the identification time and lay the foundation for real-time recognition. The system can reduce the working pressure of security personnel and protect the personal and property safety of people, so it has certain application value and popularization value.
【學(xué)位授予單位】：貴州民族大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2017
【分類號】：TP391.41

【參考文獻】

相關(guān)期刊論文 前10條

1 楊鳴鳴;;基于嵌入式系統(tǒng)的人臉識別算法研究及其優(yōu)化[J];微型機與應(yīng)用;2016年19期

2 牟琦;汪堂洪;李占利;;改進的Gabor-CRC人臉識別算法[J];計算機工程與設(shè)計;2016年10期

3 王偉凝;王勵;趙明權(quán);蔡成加;師婷婷;徐向民;;基于并行深度卷積神經(jīng)網(wǎng)絡(luò)的圖像美感分類[J];自動化學(xué)報;2016年06期

4 張文達;許悅雷;倪嘉成;馬時平;史鶴歡;;基于多尺度分塊卷積神經(jīng)網(wǎng)絡(luò)的圖像目標(biāo)識別算法[J];計算機應(yīng)用;2016年04期

5 王堅;張媛媛;柴艷妹;;基于神經(jīng)網(wǎng)絡(luò)的快速核子空間人臉識別算法研究[J];計算機科學(xué);2015年S2期

6 王亞南;蘇劍波;;基于圖像合成的多姿態(tài)人臉圖像識別方法[J];模式識別與人工智能;2015年09期

7 芮挺;費建超;周怞;方虎生;朱經(jīng)緯;;基于深度卷積神經(jīng)網(wǎng)絡(luò)的行人檢測[J];計算機工程與應(yīng)用;2016年13期

8 周全;魏昕;陳建新;鄭寶玉;;一種基于稠密SIFT特征對齊的稀疏表達人臉識別算法[J];電子與信息學(xué)報;2015年08期

9 周振華;;AdaBoost人臉檢測定點型優(yōu)化算法[J];電子科技大學(xué)學(xué)報;2015年04期

10 劉琮;許維勝;吳啟迪;;時空域深度卷積神經(jīng)網(wǎng)絡(luò)及其在行為識別上的應(yīng)用[J];計算機科學(xué);2015年07期

相關(guān)碩士學(xué)位論文 前3條

1 何西麟;基于深度學(xué)習(xí)的手寫體字符識別研究與實現(xiàn)[D];中山大學(xué);2015年

2 王強;基于CNN的字符識別方法研究[D];天津師范大學(xué);2014年

3 劉婭;小型出入口監(jiān)控系統(tǒng)中身份驗證設(shè)計[D];河北科技大學(xué);2010年

，

本文編號：1586934