【摘要】：卷積神經(jīng)網(wǎng)絡(luò)是將人工神經(jīng)網(wǎng)絡(luò)技術(shù)與深度學(xué)習(xí)方法相結(jié)合的一種新型人工神經(jīng)網(wǎng)絡(luò)模型,因?yàn)樵诖笮蛨D像處理中具有出色的表現(xiàn),使得其在計(jì)算機(jī)視覺(jué)領(lǐng)域得到了廣泛的應(yīng)用。非限定性條件下的人臉識(shí)別難度較大,應(yīng)用傳統(tǒng)的人臉識(shí)別方法難以獲得令人滿意的結(jié)果。通過(guò)設(shè)計(jì)一個(gè)深度卷積神經(jīng)網(wǎng)絡(luò),對(duì)人臉樣本進(jìn)行特征學(xué)習(xí),可以使在非限定條件下的人臉識(shí)別具有較高的準(zhǔn)確率。本文的目的在于設(shè)計(jì)一種具有較高魯棒性的深度卷積神經(jīng)網(wǎng)絡(luò),用于非限定性條件下的人臉識(shí)別,在保證準(zhǔn)確率的前提下提高系統(tǒng)的運(yùn)算效率。本文的主要內(nèi)容包括:首先,給出了卷積神經(jīng)網(wǎng)絡(luò)的理論推導(dǎo)。對(duì)在手寫(xiě)字符識(shí)別領(lǐng)域應(yīng)用廣泛的LeNet-5網(wǎng)絡(luò)模型結(jié)構(gòu)進(jìn)行了說(shuō)明,然后介紹了一種改進(jìn)的LeNet-5網(wǎng)絡(luò)用于普通的人臉識(shí)別。通過(guò)實(shí)驗(yàn)說(shuō)明了改進(jìn)后的LeNet-5可以在普通的人臉識(shí)別方面取得較好的成果。其次,對(duì)在非限定性人臉識(shí)別領(lǐng)域取得較好測(cè)試效果的VGG-19網(wǎng)絡(luò)的結(jié)構(gòu)特點(diǎn)與參數(shù)配置進(jìn)行了分析,提出了一種針對(duì)VGG-19網(wǎng)絡(luò)模型的改進(jìn)方法。通過(guò)對(duì)原始的VGG-19網(wǎng)絡(luò)結(jié)構(gòu)進(jìn)行修改,合理減少網(wǎng)絡(luò)訓(xùn)練參數(shù),在保證一定準(zhǔn)確率的情況下,降低了原網(wǎng)絡(luò)較為苛刻的硬件要求,提高了網(wǎng)絡(luò)的運(yùn)算效率。然后,在人臉數(shù)據(jù)庫(kù)FaceScrub上對(duì)新設(shè)計(jì)的卷積神經(jīng)網(wǎng)絡(luò)模型進(jìn)行了訓(xùn)練與測(cè)試,取得了較好的識(shí)別效果。最后,對(duì)實(shí)驗(yàn)結(jié)果進(jìn)行了分析,指出了該網(wǎng)絡(luò)的優(yōu)點(diǎn)及缺點(diǎn)。
[Abstract]:Convolution neural network is a new artificial neural network model which combines artificial neural network technology and deep learning method. Because it has excellent performance in the large image processing, it has been widely used in the field of computer vision. The face recognition under non restrictive condition is difficult, and the traditional face is applied. The recognition method is difficult to obtain a satisfactory result. By designing a deep convolution neural network, the feature learning of the face samples can make the face recognition with non finite conditions have a higher accuracy. The purpose of this paper is to design a kind of deep convolution neural network with high robustness for non restrictive conditions. The main contents of this paper are as follows: first, the theoretical derivation of the convolution neural network is given. A wide application of LeNet-5 network model structure in the field of handwritten character recognition is described, and then an improved LeNet-5 network is introduced for common use. Face recognition. The experimental results show that the improved LeNet-5 can achieve good results in common face recognition. Secondly, the structural features and parameter configuration of VGG-19 network with better test results in the field of non restrictive face recognition are analyzed, and an improved method for VGG-19 network model is proposed. To modify the original VGG-19 network structure, reduce the network training parameters reasonably, reduce the hard hardware requirements of the original network and improve the computing efficiency of the network. Then, the new design of the convolution neural network model is trained and tested on the face database FaceScrub. Finally, the experimental results are analyzed, and the advantages and disadvantages of the network are pointed out.
【學(xué)位授予單位】：哈爾濱工業(yè)大學(xué)
【學(xué)位級(jí)別】：碩士
【學(xué)位授予年份】：2017
【分類號(hào)】：TP391.41;TP183

【參考文獻(xiàn)】

相關(guān)期刊論文 前3條

1 黃偉;李曉玲;;基于大數(shù)據(jù)和多模態(tài)智能技術(shù)的計(jì)算機(jī)視覺(jué)實(shí)驗(yàn)設(shè)計(jì)[J];實(shí)驗(yàn)技術(shù)與管理;2016年09期

2 陳耀丹;王連明;;基于卷積神經(jīng)網(wǎng)絡(luò)的人臉識(shí)別方法[J];東北師大學(xué)報(bào)(自然科學(xué)版);2016年02期

3 徐霄;陳陽(yáng);張飛云;喬宇;;基于深度反卷積網(wǎng)絡(luò)的虹膜定位方法研究[J];集成技術(shù);2016年01期

，

本文編號(hào)：2143871