【摘要】：近年來,人臉表情識(shí)別在社交網(wǎng)絡(luò)和人機(jī)交互領(lǐng)域越來越引起學(xué)術(shù)界的重視和關(guān)注,并且已經(jīng)取得了一系列的成果�，F(xiàn)有數(shù)據(jù)庫(kù)中人臉大多角度端正、分辨率高并且環(huán)境光照良好,并且現(xiàn)有的算法均基于以上數(shù)據(jù)庫(kù)設(shè)計(jì)。而真實(shí)世界中的人臉更具有多變性,因此現(xiàn)有的算法很難滿足于實(shí)際需求。為測(cè)試現(xiàn)有算法的性能,本文探索了一些影響真實(shí)生活場(chǎng)景中笑臉檢測(cè)的因素,包括光照預(yù)處理方法、對(duì)齊、圖像尺寸、特征以及SVM分類器核的選取。根據(jù)數(shù)據(jù)本文驗(yàn)證了現(xiàn)有光照處理方法的局限性,對(duì)齊作用的實(shí)用性以分類器的核的性能等。同時(shí)為了驗(yàn)證多表情分類問題,本文通過互聯(lián)網(wǎng)搜集并建立了一個(gè)有將近3萬張人臉圖像的數(shù)據(jù)庫(kù),Real-world Affective Face Database(RAF-DB),其中每一張人臉圖像的標(biāo)簽都是被大概40位志愿者進(jìn)行獨(dú)立標(biāo)注。為了測(cè)試本文建立的RAF-DB數(shù)據(jù)庫(kù)的性能,引入了CK數(shù)據(jù)庫(kù)進(jìn)行對(duì)比,通過交叉訓(xùn)練,實(shí)驗(yàn)結(jié)果表明經(jīng)過RAF-DB訓(xùn)練在CK數(shù)據(jù)庫(kù)測(cè)試的數(shù)據(jù)結(jié)果的識(shí)別率基本上都高十在CK數(shù)據(jù)庫(kù)訓(xùn)練在RAF-DB數(shù)據(jù)庫(kù)測(cè)試的結(jié)果。搜集的數(shù)據(jù)庫(kù)表明人臉表情識(shí)別任務(wù)是一個(gè)典型的非均勻多標(biāo)簽的分類問題,為了解決上述問題,本文在訓(xùn)練時(shí),通過上采樣問題進(jìn)行了數(shù)據(jù)重構(gòu),同時(shí)也探究了多標(biāo)簽的影響,試驗(yàn)結(jié)果表明,這對(duì)識(shí)別率的提高非常明顯。在特征選取方面,除了利用比較成熟的人臉表情特征(HOG,Gabor,LBP)作對(duì)比外,還引入了深度學(xué)習(xí)的特征。對(duì)于不同數(shù)據(jù)庫(kù)以及分類任務(wù),不同的SVM核的性能差異也非常明顯,因此本文分類器的訓(xùn)練采用了多核SVM分類器,包括線性核、高斯核以及局部線性核(OCC)。試驗(yàn)結(jié)果表明,多核SVM在表情分類問題上具有更強(qiáng)的穩(wěn)定性和更高的準(zhǔn)確率。
[Abstract]:In recent years, facial expression recognition has attracted more and more attention in the field of social network and human-computer interaction, and has made a series of achievements. In the existing database, the human face is large and multi-angle correct, the resolution is high and the environment illumination is good, and the existing algorithms are based on the above database design. In the real world, human faces are more variable, so the existing algorithms are difficult to meet the actual needs. In order to test the performance of the existing algorithms, this paper explores some factors that affect the detection of smiling faces in real life scenes, including illumination preprocessing, alignment, image size, features and the selection of SVM classifier cores. According to the data, this paper verifies the limitation of existing illumination processing methods and the practicability of alignment to the performance of classifier kernel. At the same time, in order to verify the problem of multi-expression classification, this paper collects and builds a database of nearly 30, 000 face images via the Internet, Real-world Affective Face Database (RAF-DB). Each face image was labeled independently by about 40 volunteers. In order to test the performance of the RAF-DB database established in this paper, the CK database is introduced and compared. The experimental results show that the recognition rate of the data tested in the CK database after RAF-DB training is almost higher than that in the CK database training in the RAF-DB database. The database collected shows that the task of facial expression recognition is a typical non-uniform multi-label classification problem. At the same time, the influence of multi-label is also discussed. The experimental results show that the recognition rate is improved obviously. In feature selection, in addition to using more mature facial expression features (HOG,Gabor,LBP) for comparison, in-depth learning features are also introduced. For different databases and classification tasks, the performance of different SVM kernels is also very different. Therefore, the training of the classifier in this paper uses multi-core SVM classifier, including linear kernel, Gao Si kernel and local linear kernel (OCC). The experimental results show that multicore SVM is more stable and accurate in facial expression classification.
【學(xué)位授予單位】：北京郵電大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2016
【分類號(hào)】：TP391.41

【參考文獻(xiàn)】

相關(guān)碩士學(xué)位論文 前1條

1 孫雯玉;人臉表情識(shí)別算法研究[D];北京交通大學(xué);2006年

，

本文編號(hào)：2353939