【作者】 张勇；

【导师】 牛常勇；

【作者基本信息】 郑州大学 ， 软件工程， 2015， 硕士

【摘要】 人脸和车牌的检测问题·真是目标检测领域的热门问题。基于人脸识别和车牌识别的应用在现实生活中发挥的作用越来越重要,尤其随着互联网科技和智能化技术的发展,人们在数据挖掘、安保监控、智能交通等相关智能系统的需求也越来越强烈。在现实中,各种各样的图像采集系统的逐渐普及,产生了海量的午牌、人脸图像信息。这些采集到的图像往往来自不同的环境和系统,导致了数据没有统一的规范格式,使得后续的信息提取和挖掘面临很多困难。同时,不受限的环境使得定协面临着许多复杂的影响因素,诸如不均匀光照、多姿态、运动模糊、低分辨率、装饰物遮挡等问题。堪管对于它们的研究起步较早,但是当前的检测系统都对环境有着一定的约束限制。同时,车牌、人脸检测问题有着许多共同之处,因此共同研究车牌和人脸的检测问题就具有重要的意义和现实价值。初期的对于车牌和人脸检测,都尝试了采用颜色、几何轮廓、纹理等视觉特征进行目标的定位。分析这些方法可以清晰的看到这些方法存在各自的优缺点和适用场景。在诸如固定场所的监控系统中,数据往往具有高分辨率、场景单一、视角固定等特点。在这种受限的特定环境,我们使用的人工设计特征的方法,可以高效且很好的解决了这两类问题。然而,存诸如室外监控等系统中,数据存在运动模糊、装饰物遮挡、环境下扰等因素。在这种自然环境下,单一、组合使用人工设计特征方法性能都严重受到了影响。因此增强车牌、人脸检测算法对不同场景、干扰因素的鲁棒性,是一个非常值得研究和解决的技术问题。20世纪90年代至今,随着人类对大脑认知的深入,基于样本统计的机器学习方法从提出之后就得到了充分的发展和完善。这种基于统计的机器学习方法能从样小数据中挖掘出中富的特征表示,使得存很多方湎较于人工设计特征都表现了巨大的优越性。这种方法极大的激发了人们对机器学习理论和应用研究的热情和信心。神经网络是机器学习优秀算法之一,是通过模拟人脑视觉神经系统的行为特征,进行分布式并行信息处理的算法数学模型,使用反向传播算法(Back Propagation)训练的人工神经网络模型很好的解决了MNIST分类问题。卷积神经网络是人工神经网络的一种。这种多阶段全局训练模型,通过样本训练自主的从原始输入中学习平移不变性特征,基于滑动窗口的技术使得算法可以用较小的代价完成图像扫描,因此在目标检测领域得到了广泛使用。2006年之后,随着深度学习概念的提出和研究的发展,尤其在语音识别、图像分类等领域,深度学习模型很好的解决了浅层学习模型难以逾越的障碍,极大的推动了机器学习的新浪潮—深度学习。本文荩于深度学习思想提出使用深度卷积神经网络(Deep Convolutional Neural Networks)模型解决复杂环境下人脸和车牌的检测问题更多还原

【Abstract】 Detection of face and the license has been a hot issue in the field of target detection. Based on the application and effect of face recognition and license plate recognition play in real life is more important, especially with the development of Internet technology and intelligent technology, people in data mining, security monitoring, intelligent transportation and other related intelligent systems have increasingly strong demand. In reality, all kinds of image acquisition system has been popular, produced a huge amount of license plate, the face image information. The collected images often come from different environments and system, resulting in no uniform format of data, so that the subsequent information extraction and mining face many difficulties. At the same time, unrestricted environment makes the positioning is faced with many complicated factors, such as uneven illumination, pose, motion blur, low resolution, decorative objects occlusions. Although their research started earlier, but the current detection system for a certain environmental constraints. At the same time, the license plate, the problem of face detection has a lot in common, so the research of license plate and the face detection problem is of great significance and practical value.Early detection of license plate and the face, try positioning using visual features of color, texture, geometry contour of targets. The analysis of these methods can clearly see all these methods have their respective advantages and disadvantages and the applicable scenes. Monitoring system in fixed locations, data often has the advantages of high resolution, single fixed characteristics, from the perspective of the scene. In the specific environment of this limited, we use the method of artificial design features, can be highly efficient and well solve the problem of the two kind. However, such as outdoor surveillance system, there is motion blur, decoration, environment factors such as occlusion interference data. In the natural environment, the use of artificial design features a single, combined method of performance are seriously affected by the. Therefore, to enhance the robustness of license plate, the face detection algorithm for different scenarios, interference factors, is a very worthy of study and solve technical problems.Since twentieth Century 90 years, along with the human brain cognitive in-depth, the statistical machine learning method based on the proposed after it has been fully developed and perfected. The statistical based machine learning methods to dig out the rich characteristics of the sample data from said that in many aspects, compared with manual design features have shown great superiority. This method greatly stimulated the research on theory and application of machine learning enthusiasm and confidence. Neural network is one of the best machine learning algorithm, by simulating the human visual characteristics of the nervous system, the distributed parallel processing algorithm of the mathematical model, using the back propagation algorithm(Back Propagation) is the MNIST classification problem of a good solution for training artificial neural network model. Convolutional neural network is an artificial neural network. The global training model, through sample training autonomous learning from the original input shift invariant feature, the sliding window technology makes the algorithm can accomplish the image scanning with low cost based, so in the field of target detection has been widely used. After 2006, with the depth of learning concept and the development of the research, especially in speech recognition, image classification and other fields, the deep learning model is a good solution to the shallow learning model of insurmountable obstacles, greatly promoted the new wave, depth of machine learning. In this paper, the depth of learning is proposed based on neural network using convolutional(Deep Convolutional Neural Networks) model to solve the problem of face detection in complex environment and vehicle license plate and a good performance on the set of the experimental data obtained.更多还原