本文選題：微表情識(shí)別　切入點(diǎn)：3-d卷積神經(jīng)網(wǎng)絡(luò)　出處：《溫州大學(xué)》2017年碩士論文

【摘要】：短時(shí)間人們表情的變化,也叫微表情,心理學(xué)在這方面的研究已經(jīng)很早就開(kāi)始了,近年來(lái),有關(guān)利用機(jī)器學(xué)習(xí)的方法來(lái)對(duì)微表情進(jìn)行研究的學(xué)者越來(lái)越多,是當(dāng)前研究的一個(gè)熱門(mén)方向。在這幾年的微表情識(shí)別的研究中,有幾個(gè)團(tuán)隊(duì)為微表情研究建立了數(shù)據(jù)集供其他研究者使用,并且提出了一些算法來(lái)解決微表情識(shí)別的問(wèn)題。自從2012年卷積神經(jīng)網(wǎng)絡(luò)在ImageNet的比賽上取得了重大突破以后,基于卷積神經(jīng)網(wǎng)絡(luò)的深度學(xué)習(xí)方法在圖像識(shí)別領(lǐng)域取得了越來(lái)越好的結(jié)果。本人在做微表情識(shí)別的研究中主要做了兩種方法:1、第一種方法是基于卷積神經(jīng)網(wǎng)絡(luò)的,主要采用了基于光流的3-d CNN的網(wǎng)絡(luò)結(jié)構(gòu),將最簡(jiǎn)單的VggNet網(wǎng)絡(luò)加上微表情視頻時(shí)間序列的信息構(gòu)成3維的信息輸入,同時(shí)將X和Y方向的3-d光流場(chǎng)信息與原始灰度的3-d信息經(jīng)過(guò)三個(gè)通道的3-d CNN處理,最后將三部分信息組合并分類。2、第二種方法是利用了積分投影和LSTM來(lái)對(duì)微表情進(jìn)行識(shí)別�，F(xiàn)有的微表情識(shí)別研究主要是利用基于局部二值模式(LBP)改進(jìn)的算法并結(jié)合支持向量機(jī)(SVM)來(lái)識(shí)別。最近,積分投影開(kāi)始應(yīng)用于人臉識(shí)別領(lǐng)域。長(zhǎng)短時(shí)記憶網(wǎng)絡(luò)(LSTM)作為循環(huán)神經(jīng)網(wǎng)絡(luò),可以用來(lái)處理時(shí)序數(shù)據(jù)。因此提出了結(jié)合積分投影和LSTM的模型(LSTM-IP),在最新的微表情數(shù)據(jù)庫(kù)CASME Ⅱ中進(jìn)行實(shí)驗(yàn)。通過(guò)積分投影得到水平和垂直投影向量作為L(zhǎng)STM輸入并分類,同時(shí)采用防止過(guò)擬合技術(shù)。實(shí).驗(yàn)結(jié)果表明,LSTM-IP算法模型取得了比以前的方法更好的精度。
[Abstract]:The changes in people's expressions in a short period of time, also known as microexpressions, have long been the subject of psychological research. In recent years, more and more scholars have studied microexpressions using machine learning methods. In recent years of research on microfacial expression recognition, several teams have created data sets for microfacial expression studies for use by other researchers. And some algorithms are put forward to solve the problem of microfacial expression recognition. Since 2012, the convolutional neural network has made a great breakthrough in the ImageNet competition. The depth learning method based on convolution neural network has obtained more and more good results in the field of image recognition. In the research of micro expression recognition, I have mainly done two methods: 1. The first method is based on convolutional neural network. The network structure of 3-d CNN based on optical flow is mainly adopted. The simplest VggNet network and the information of microfacial video time series are used to form 3D information input. At the same time, the 3-d optical flow field information in X and Y directions and the 3-d information in the original gray level are processed by 3-d CNN with three channels. Finally, the three parts of information are combined and classified. 2. The second method is to use integral projection and LSTM to recognize micro-expression. The existing research on micro-expression recognition is mainly based on the improved algorithm based on local binary pattern. Support vector machine (SVM) to identify. Recently, Integral projection has been applied to face recognition. LSTM (long and short time memory Network) is used as a cyclic neural network. It can be used to deal with time series data. Therefore, a model combining integral projection and LSTM is proposed to carry out experiments in the latest microfacial expression database CASME 鈪，

本文編號(hào)：1667067