本文選題：深度學(xué)習(xí)　切入點：人臉識別　出處：《大連理工大學(xué)》2013年碩士論文　論文類型：學(xué)位論文

【摘要】：在實際應(yīng)用中,如監(jiān)控系統(tǒng)中,采集到的人臉圖像往往是具有多種姿態(tài)變化的,并且圖像分辨率偏低。受姿態(tài)變化和分辨率低的影響,造成的人臉圖像識別性能的迅速下降,而姿態(tài)變化是人臉識別中一個最為突出的難題。姿態(tài)變化將非線性因素引入了人臉識別中。而現(xiàn)有的一些機(jī)器學(xué)習(xí)方法(如只有一個隱層的神經(jīng)網(wǎng)絡(luò)、核回歸,支撐向量機(jī)等)大都使用淺層結(jié)構(gòu)。心理學(xué)研究表明對于有限數(shù)量的樣本和計算單元,淺層結(jié)構(gòu)難以有效地表示復(fù)雜函數(shù),并且對于復(fù)雜分類問題表現(xiàn)性能及泛化能力針均有明顯的不足,尤其當(dāng)目標(biāo)對象具有豐富的含義。深度學(xué)習(xí)可通過學(xué)習(xí)一種深層非線性網(wǎng)絡(luò)結(jié)構(gòu),實現(xiàn)復(fù)雜函數(shù)逼近,表征輸入數(shù)據(jù)分布式表示,并體現(xiàn)了它對于輸入樣本數(shù)據(jù)的強(qiáng)大的本質(zhì)特征的抽取能力。因此本文運(yùn)用深度神經(jīng)網(wǎng)絡(luò)的方法克服姿態(tài)變量和圖像分辨率的影響,提出了一種多姿態(tài)的人臉超分辨識別算法并在實驗數(shù)據(jù)集上獲得了良好的性能表現(xiàn)。 另外本文利用深度信念網(wǎng)絡(luò)探索正面人臉圖像和側(cè)面人臉圖像的映射,方法放松了深度信念網(wǎng)絡(luò)的輸入也輸出之間絕對相等,而只是保證其高層含義上的相等。實驗表明了使用深度信念網(wǎng)絡(luò)可以學(xué)習(xí)到側(cè)面人臉圖像到正面人臉圖像的一個全局映射,但是丟失了個體細(xì)節(jié)差異。本文還提出了基于深度網(wǎng)絡(luò)保持姿態(tài)鄰域進(jìn)行姿態(tài)分類的方法,在學(xué)習(xí)過程中,我們保證了同一個姿態(tài)下的人臉圖像應(yīng)該與同一姿態(tài)下的多張圖像互為鄰居。實驗證明了,我們的方法在用于姿態(tài)分類具有非常良好的性能,但是也發(fā)現(xiàn)學(xué)習(xí)過程中,那些與區(qū)別個體的信息逐漸丟失了,這也導(dǎo)致了直接運(yùn)用非線性近鄰元分析的特征的人臉識別的性能不佳。 本文是一篇基于深度學(xué)習(xí)在人臉識別姿態(tài)和分辨率問題上的研究,此外,本文還探索了深度信念網(wǎng)絡(luò)在人臉姿態(tài)處理中的潛在應(yīng)用,如姿態(tài)映射和姿態(tài)分類。
[Abstract]:In practical applications, such as the monitoring system, the collected face images often have a variety of pose changes, and the image resolution is low, which is affected by the change of the pose and the low resolution, resulting in a rapid decline in the performance of face image recognition. Pose change is one of the most prominent problems in face recognition. Attitude change introduces nonlinear factors into face recognition. Some existing machine learning methods (such as neural network with only one hidden layer, kernel regression, etc.). Psychological studies show that for a limited number of samples and computing units, shallow structures are difficult to represent complex functions effectively. Moreover, the performance and generalization ability of complex classification problems are obviously inadequate, especially if the target object has rich meanings. Depth learning can achieve complex function approximation by learning a deep nonlinear network structure. The distributed representation of input data is characterized by its strong ability to extract essential features of input sample data. Therefore, the influence of attitude variables and image resolution is overcome by using the method of depth neural network. A multi-pose super-resolution face recognition algorithm is proposed and a good performance is obtained on the experimental data set. In addition, this paper uses depth belief network to explore the mapping of frontal face image and side face image. The method relaxes that the input and output of depth belief network are absolutely equal. The experiment shows that the depth belief network can be used to learn a global mapping from the side face image to the frontal face image. However, the differences of individual details are lost. In this paper, we also propose a method of attitude classification based on depth network to maintain attitude neighborhood, in the process of learning, We ensure that the face image in the same pose should be neighbors with multiple images in the same pose. Experiments show that our method has a very good performance in attitude classification, but it is also found in the process of learning. The loss of information from the individual leads to poor performance of face recognition based on the features of nonlinear nearest neighbor element analysis (NNEM). This paper is a study on face recognition pose and resolution based on depth learning. In addition, this paper also explores the potential applications of depth belief network in face pose processing, such as pose mapping and pose classification.
【學(xué)位授予單位】：大連理工大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2013
【分類號】：TP391.41

【參考文獻(xiàn)】

相關(guān)期刊論文 前1條

1 孫冬梅,裘正定;生物特征識別技術(shù)綜述[J];電子學(xué)報;2001年S1期

相關(guān)博士學(xué)位論文 前1條

1 楊健;線性投影分析的理論與算法及其在特征抽取中的應(yīng)用研究[D];南京理工大學(xué);2002年

，

本文編號：1603075