【作者】 林忠华；

【导师】 苑玮琦；

【作者基本信息】 沈阳工业大学 ， 电工理论与新技术， 2008， 博士

【摘要】 随着现代化技术的不断出现,传统方法已经不能满足当今社会对身份识别的要求。这样,就使得生物特征识别技术得以迅速发展起来。生物特征识别具有传统方法所无法比拟的高准确性、高安全性、高可靠性等优点,能够满足社会各个领域的需求。虹膜身份识别是一种新兴的生物身份识别技术,由于其具有唯一性、稳定性、可采集性、非侵犯性等优点而逐步受到人们的重视。与脸像、声音、指纹等身份鉴别方法相比,虹膜具有更高的准确性。近些年来虹膜识别技术在研究和应用方面都得到了长足的进步,并表现出了广阔的前景和市场。虹膜识别系统主要包括了虹膜图像采集、虹膜图像预处理、虹膜图像特征提取、模式匹配与识别。它的研究主要涉及到了计算机视觉、数字图像处理、小波理论、模式识别等众多学科领域。本文在总结目前虹膜识别技术研究进展的基础上,对虹膜识别中的虹膜定位、噪声处理、归一化、特征提取及编码和分类决策问题展开探讨,提出了自己的一些改进方法并进行实验对比,对进一步研究虹膜识别技术有一定的借鉴作用。在虹膜定位方面,提出一种基于人眼结构特征的定位方法。首先,通过邻域灰度平均值去除光斑,接着通过投影法和瞳孔中心检测算子来找瞳孔内一点,然后通过边缘检测算子结合投票的方法找内外四个边界点准确定位虹膜内外边界。对2655幅人眼图像进行虹膜定位实验的结果证明,该方法能够有效地解决虹膜图像平移问题,并避免了反射光斑、眼睑、眼睫毛等噪声对定位的影响。与经典方法相比,该方法在保证定位成功率的同时,定位时间大大减少。在人眼自然睁开状态下存在噪声,因此需要进行去噪操作:通过抛物线拟合的方法去除上、下眼睑;利用邻域灰度均值法去除光斑;去除眼睫毛时先用局部灰度极小值的方法找到所有候选眼睫毛点,然后利用眼睫毛的起始信息、方向信息和长度信息去除伪眼睫毛点。在归一化方面,将环形虹膜图像通过坐标变换转变为矩形,该方法能够有效地解决虹膜图像缩放问题,同时进行内外圆心不一致的补偿操作,并确定有效虹膜区域,便于进行后续的操作。在特征提取和编码方面,在有效虹膜区域内,采用基于小波变换系数的特征提取方法,并提出基于局部灰度极小值、基于纹理特征点匹配,基于多方向纹理边缘检测三种结构特征提取方法。基于小波变换系数的特征提取方法采用墨西哥草帽小波和Morlet小波,对不同尺度下的小波变换系数分布进行考察,通过设置不同的尺度,可以得到不同频率下的纹理特征,并分析哪些尺度下得到的虹膜纹理比例要高一些。与传统的小波变换方法相比,本文提出的基于小波变换系数的方法更具有优势,不仅算法的复杂度比较低,而且算法的识别率高;基于局部灰度极小值的特征提取方法利用了虹膜纹理的位置信息和灰度信息,却抛弃了纹理的大小、方向、相关性等其他结构特征,因此提取出来的可区分性特征仍然不能保证使虹膜识别的准确性得到大幅度的提高;基于纹理特征点匹配的特征提取方法虽然利用了纹理的大小与方向信息,但是选取的方向信息少,未考虑纹理之间的相关性,使得识别准确性还有可提高的空间;基于多方向纹理边缘检测的特征提取方法利用了纹理的位置、灰度、大小、方向、相关性等多种结构特征,因此提取出来的可区分性特征使虹膜识别的准确性得到大幅度的提高。在模式匹配与识别方面,采用自行设定的计算距离的方法来计算登陆虹膜码与注册虹膜码之间的距离,通过多次平移匹配来解决图像旋转的问题,通过阈值法进行分类。最后,在已有的图库上,对基于Gabor滤波器的方法进行了实验,并将本文提出的方法与基于Gabor滤波器的方法进行了比较和分析。更多还原

【Abstract】 With the constant development of modern technologies, traditional methods of personal recognition can not satisfy various demands of the present society. Under this condition, the personal recognition based on the biometric begins to develop rapidly. The rising biometrics has a lot of advantages which are not involved in traditional methods, such as high accuracy, high security, high reliability, and so on. Biometrics also satisfies many requirements of each field in our society. Iris recognition is an emerging biometric technology. For its exclusive, stability, collectible and unforced, iris recognition is being more and more regarded by people. In recent years, iris recognition has made progress in technology research and application, and has a wide prospect and market. Compared with face, voice, fingerprint and other biometric technology, iris recognition has higher precision.An iris recognition system includes iris imaging, iris preprocessing, feature extraction, pattern match and classifying. Its research aspects involve many subjects such as computer vision, digital image processing, wavelet theory, pattern recognition. Based on the recent advancements in iris recognition, this article makes discussion about iris localization, noise processing, normalization, feature extraction, code and classifying decision-making, and presents some improving methods and the experimental results are also described. These use for references for more research on the iris recognition technology.In iris localization, a method of iris localization based on the human eye structure is presented. Firstly, it removes the facula by the average gray of adjacent area and finds a point in the pupil by the projection and pupil-center detection operator, then finds four boundary points in inner and outer boundary respectively by edge detection operator or combining voting method to localize the inner and outer boundaries. Experimental results on 2655 eye images demonstrate that the proposed method of iris localization can effectively resolve the problems of iris translation and it can not be affected by facula, eyelid and eyelashes. Compared with two classical methods, the proposed method ensures the probability of successful localization and reduces the time consumedly.There is much noise in the state of eyes opening unforcedly, so the noise processing is necessary, the upper and lower eyelids are removed by parabola fitting, the facular is removed by neighboring gray average, the eyelashes are removed by finding all the candidate eyelashes by local gray minimum firstly, then the false eyelashes are removed by the beginning location, directional and length information of the eyelashes.In normalization, the annular iris image is normalized to rectangular image. It can resolve the problems of iris zoom and solve the disaccord of inner and outer center, and ensure the effective iris area for the next operations.In feature extraction and code, many feature extraction methods are presented such as based on wavelet transform coefficients, local gray minimum, texture feature points matching, multidirectional texture edge detection. The method based on the wavelet transform coefficients adopts Marr and Morlet wavelets to extract feature based on the wavelet transform coefficients, reviews the coefficients distribution of different scales. It can get feature of different frequency by setting different scale and analyze which iris texture proportion is higher under which scale. This method has more advantages, lower complication and higher recognition rate comparing with traditional wavelet transform methods. The method based on local gray minimum uses the position and gray information of iris texture, but discards other structure characteristic such as size, direction and relativity of texture, so the partial feature can not ensure to increase much more correctness. Although the method based on texture feature points matching uses the size and direction information, the direction information chosen is less and discards the relativity of texture, so there is much more elevated space. The method based on the multidirectional edge detection uses the position, gray, size, direction and relativity of texture, so the dividing features ensure to increase the precision much more.In pattern match and recognition, it calculates the distance between enroll iris code and register iris code by distance designed by ourselves, solves the rotary problems by many transferring matches along horizontal directions, and classifies by threshold.Finally the performance of the presented methods and the method based on the gabor filter is analyzed and compared in the V3 Interval database.更多还原