摘要目前的黑市上流通着千万级的姓名、身份证、照片等关键信息，用户的身份 信息甚至是密码都已经把握在一些不法分子手里，所以传统的密码已经不再安全。 人脸识别是目前技术水平下最容易实现的识别方式，它具有硬件要求低、识别率 高、不易伪造等优点。本文以局部二值模式和深度信念网络相结合实现人脸的快 速识别，可以保证在少量样本情况下准确识别人脸。因此本文算法可以应用于实 际的门禁系统、支付系统中，实现高安全度的身份验证。83387

本文主要研究局部二值模式对人脸纹理信息的提取，以及深度信念网络对局 部二值模式提取出来的纹理特征进行模式匹配。本文算法在预处理阶段先使用肤 色检测和模板匹配的快速人脸搜索算法将人脸图片从背景中裁剪出来，大大减少 了背景对人脸识别的影响。本文算法包括：人脸的定位、人脸裁剪与归一化、局 部二值模式提取纹理特征、深度信念网络识别人脸，并通过算法的实验找到合适 的深度信念网络隐藏单元数。

本文算法在 ORL 人脸库上当每个人选用 5 张共选用 40 个人的图片作为训练 样本，样本大小为 92*112 像素。当隐藏单元数取 100 时识别正确率到达 95。83%， 训练用时 78 秒；隐藏单元数取 200 时识别正确率可以到达 98。33%，训练用时 170 秒。可以看出当隐藏单元数增加时，识别的正确率上升，同时训练用时也增加。 在实际应用中，可以在保证正确率达到要求时，选取适当的隐藏单元数来减少训 练用时，加快算法运行速度。本文找出单元数为 100 时，可以达到较高的正确率 同时训练用时也比较小，是比较适合的隐藏单元数。

毕业论文关键词：人脸识别；局部二值模式；深度信念网络；人脸检测；肤色检测

Abstract There are tens of millions of names, identity cards, photographs and other key information on the black market circulation nowdays。 The user's identity information and even the password has been held in the hands of criminals。 So the traditional password is no longer safe。 Face recognition is the most easily realized recognition method under the current technical level, it has the advantages of low hardware requirements, high recognition rate, is not easy to forge and so on。In this paper, based on the local binary pattern and deep belief network, face recognition is realized。 It can be realized in a small number of samples to accurately identify the face。 Therefore, the algorithm can be applied to the actual access control system, payment system, to achieve a high degree of security authentication。

In this paper, we mainly study the extraction of facial texture information by local binary pattern ,and the pattern matching of the texture features extracted from the local binary pattern by the deep belief network。 In this paper, we first use skin color detection and template matching in the preprocessing stage to cut the face images from the background, which greatly reduces the impact of the background on the face recognition。 In this paper, the algorithm includes: the location of the human face, the face cut and the normalization, the local binary pattern to extract the texture feature, the deep believe network to recognize the human face, and and through the algorithm of the experiment to find the appropriate number of hidden units of deep believe network 。

In this paper, the algorithm in the ORL face database for each person to choose 5 as a training sample, the sample size of 92*112 pixels。 When the number of hidden units is 100, the recognition accuracy reaches 95。83%, and the training time is 78 seconds。 When the number of hidden units is 200, the recognition accuracy can reach 98。33%, and the training time is 170 seconds。 It can be seen that when the number of hidden units is increased, the correct rate of recognition is increased, and the training time is also increased。 In practical application, in ensuring the correct rate to meet  the