论文总字数:30228字
摘 要
如何让计算机像人类一样准确地识别人脸,在如今具有很重要的现实意义。最近火热的刷脸支付尤其体现了这一点,可以说人脸识别蕴含着大量的商业价值。另外,在凶手追踪、目标识别等方面也大有用武之地。由此可见,人脸识别的研究具有很重要的现实意义。人脸识别的算法有很多。主成分分析法[1]以及其衍生出的改进算法,如二维主成分分析法[2]、核主成份分析法[3]等在人脸识别上都有不错的效果。主成分分析法实现了对人脸图像的降维,对降维后的数据再进行分类会有较好的效果。尝试比较了不同分类算法的效果,线性判别分析、K-近邻、BP神经网络都能达到较高的识别率。同时自己尝试对BP神经网络进行改进,引入了弹性梯度下降法,使得每次的步长会根据具体情况实时调整,缩短了训练时间。同时加入了随机概率,有利于权值在训练过程中跳出局部极值,得到全局最优解。因为降维后的数据并不庞大,改进后的BP神经网络在权值训练上所耗的时间也并不太多,将神经网络用于人脸分类是很合适的。
人脸包含多种特征,局部二值模式[4]可以有效提取人脸图像的纹理特征,梯度方向直方图[5]则可以很好地得到图像的轮廓。通过将两种特征结合起来,实现人脸特征融合,可以较好地保留人脸有用信息。其中的局部二值模式使用等价模式,从而在一定程度上减少了数据的稀疏性。之后使用主成分分析法实现数据降维,再利用BP神经网络[6]进行分类,达到了更高的人脸识别率。为了增强人脸识别对于角度旋转的鲁棒性,人为地对人脸数据库中的图像使用双线性插值算法[7]进行旋转,这样人脸库中便会存在各种角度的人脸图像,对于提升人脸旋转下的识别率有很好的效果。
不过算法在光照变化较大、人脸有穿戴物遮挡等情况下的识别率还有待进一步提高。人有较为夸张的表情变化时,识别率也会有一定幅度下降。在这些方面还需要进一步的改善。
关键字:人脸识别、主成分分析法、BP神经网络、局部二值模式、梯度方向直方图、二维PCA
Face Recognition Research based on fusion of various features
Abstract
Making a machine distinguish one from another one just like a human being is of great importance today. Recently, paying with your face is becoming popular which indicates face recognition contains a large amount of commercial value. In addition, the face recognition can also be used in conditions like killer tracking, public safety and so on. This shows face recognition has a good practical significance. A lot of algorithms can be used in face recognition such as principal component analysis(PCA), 2DPCA、KPCA and so on. These algorithms have good results in face recognition. Principal component analysis realizes the dimension reduction of facial images and the reduction of the dimension of data will have better results in classification. We try to compare the effects of different classification algorithms. Linear discriminant analysis, K- nearest neighbors and BP neural network can achieve higher recognition rate. At the same time, we use the elastic gradient descent to improve BP neural network where the training steps will be adjusted depending on the circumstances. Random probability is also used with the intention of achieving global optimal solution instead of local optimal solution. After dimensionality reduction, the data is not large. Thus, training the improved BP neural network weights do not consume too much time. The neural network for face classification is appropriate.
Face images include a variety of features. Local binary pattern can effectively extract texture features of the face images and histogram of oriented gradient can represent image contours. By combining the two features to achieve fusion of facial features, we can retain the useful information of face images. For local binary pattern, the uniform patterns are used with which the sparsity of data can be reduced to some extent. Use principal component analysis for data dimensionality reduction, and then re-use BP neural network to classify, we can achieve a higher recognition rate. In order to enhance the robustness of face rotation, we use the bilinear interpolation algorithm to rotate the face images in database. So that face images rotated with various angles will exist in database. Based on this, the face recognition accuracy with image rotated is rather good.
The recognition accuracy is not satisfactory when illumination changes dramatically. What’s more, when putting on glasses or wearing a hat, the recognition accuracy is also not good. We still have a long way to go.
Key words: Face recognition,PCA,2DPCA,BP neural network, Local binary pattern, histogram of oriented gradient
目录
第一章 绪论 1
1.1引言 1
1.2人脸识别技术的研究现状以及发展前景 2
1.3 人脸识别过程 3
1.4论文结构安排 3
第二章 人脸预处理 5
2.1消除光照的影响 5
2.2 去噪滤波 7
第三章 多种人脸识别算法 10
3.1 主成分分析法(PCA) 10
3.2 线性判别分析(LDA) 11
3.3 二维主成分分析法(2DPCA) 13
3.4 BP神经网络 15
3.5 局部二值模式结合梯度方向直方图的分层特征融合 17
3.6图像旋转提高人脸识别率 26
第四章 实验仿真结果分析 29
4.1PCA与二维PCA的比较 29
4.2 PCA BP网络算法对比 31
4.3 LBP、HOG融合特征结合PCA算法 32
4.4旋转不变性仿真实验 34
第五章 总结与展望 35
5.1全文工作总结 35
5.2研究展望 36
致谢 37
参考文献 38
第一章 绪论
1.1引言
计算机视觉是当今计算机领域发展的一大方向。未来无论是无人驾驶、工厂自动化监测,亦或是各种类型的具有较高自主性的机器人,它们都需要一双明亮且敏锐的“眼睛”。通过对观察到的图像进行实时地分析,并具有近乎百分之百的识别率,它们可以高效率地替人类完成各种繁杂的任务。
如今在中国,最火的莫过于百度的无人驾驶汽车了。吴恩达领导的团队已经宣布将会很快在美国测试自动驾驶汽车。百度打算在2018年推出具有自动驾驶功能的穿梭车,安全起见,它只在限定区域内环形运营。这样可以有效减少突发情况的发生,保证乘客的安全。随着系统不断地自我学习,它的运营范围也会相应扩大。由此可见,机器视觉这一领域在未来必定大有可为。
剩余内容已隐藏,请支付后下载全文,论文总字数:30228字
该课题毕业论文、开题报告、外文翻译、程序设计、图纸设计等资料可联系客服协助查找;