【作者】 孟凡伟；

【导师】 赵业权；

【作者基本信息】 哈尔滨工业大学 ， 飞行器设计， 2010， 博士

【摘要】 体全息存储技术以其存储密度高、存储容量大、数据传输速率高、数据搜索时间短等优势成为一种颇具潜力的海量信息存储技术。而基于体全息存储的相关识别技术是一种新兴的光学模式识别技术。具有并行性和快速性的特点,在军事制导及目标识别等方面具有重要的研究价值及应用前景。本论文以近化学计量比Mg:Fe:LiNbO₃晶体为存储介质,在已有的理论和研究基础上,提出了若干适用于体全息相关识别的算法,推进体全息识别系统的实用化进程。目前制约体全息存储及相关识别技术的关键仍然在于获取高性能、高质量的光折变材料。同成分LiNbO₃晶体难以满足体全息存储对其光折变性能的要求,而且在强光照射下容易发生光致散射。为此本文利用顶部籽晶法,通过优化晶体生长的工艺参数（原料配比、温场等）,研制出大尺寸、组分均匀、抗光损伤能力强的近化学计量比Mg:Fe:LiNbO₃晶体。测试结果显示:晶体中的[Li]/[Nb]比在49.7⁴9.8%之间,确定了Mg在晶体中的阈值浓度为2³mol%之间,提出了掺杂离子的占位模型,晶体的头尾居里温度差在2℃以内,晶体的抗光损伤能力达到10⁴W·cm^-2,衍射效率超过60%,响应时间达到亚秒级。表明近化学计量比Mg:Fe:LiNbO₃晶体具有更好的光学均匀性和体全息存储特性,较同成分LiNbO₃晶体有显著提高,为提高体全息相关识别系统的读出图像质量打下坚实的基础。为了进一步降低体全息相关识别系统对全息图衍射效率均匀性的要求并提高识别率,采用基于特征识别向量的方法进行识别判断,该方法需要对相关峰光斑进行自动定位。为此,以邻域熵目标检测为基础,提出了一种基于邻域方差增长的相关峰光斑定位算法。首先选择大小合适的两邻域,通过计算两个邻域的方差增长数值来而不是熵增长值来进行目标检测和定位,同时解决了邻域熵进行目标检测受背景亮度影响的问题。针对1024×768像素大小的相关读出图像,算法定位比较准确,且平均耗时<4s。针对大样本问题,提出了基于自适应多尺度边缘的体全息相关识别算法,使原始图像之间的直接相关转化为特征边缘之间的相关,提高了相关峰的锐度。首先依据噪声和边缘在小波变换域的Lipschitz指数的差异性,定义多尺度边缘相关函数,对模极大值点进行检测;然后依据类内方差最小化这一准则自适应的确定二值化的双阈值,避免人为因素的干扰。以AR人脸库中图像作为原始图像,实现了1000幅全息图的存储与相关识别,对于库内图像识别准确率达到了99.50%,对于500幅库外图像都能够准确的判断为非库内图像。提出了一种加权二维Fisherface方法对训练样本进行特征提取和重构,来解决体全息相关识别所面临的小样本问题。算法首先使用二维主成分分析方法对原始数据进行压缩降维,使得压缩后的信息能够最大限度的表达原始输入信息;然后根据Fisher准则在降维后的空间进行特征提取,使投影后的模式样本在新的子空间中类间离散度和类内离散度的比值最大,即模式在该空间具有最佳的分离性,从而达到提取分类信息的效果。最后根据这些最具分离性的特征和最具表达性的特征对原始图像进行重构,将原始图像之间的直接相关转化为重构图像之间的相关。同时算法对那些偏离聚类中心的野值点在计算类均值时赋予较轻的权重,也在一定程度上提高了所提取的特征的准确性。在ORL和Yale人脸库上的相关识别实验取得的识别率较传统相关提高了约10%,进一步说明了算法的有效性。更多还原

【Abstract】 With the developing of information and computer technology, the new storage system is needed urgently. Volume holographic storage has become a potential storage means for the merits of high package density, high storage capacity, fast data transfer rates and short access time. Accordingly, the correlator based on volume holographic storage is a rising optical pattern recognition technology with inherent characteristics of parallelism and rapidity.The preparition of photorefractive materials with high performance and better quality is the key technology to promote the development of volume holographic correlation. So, near stoichiometric Mg:Fe:LiNbO₃ crystals, which is not only heavy gauge and uniform component but also high photodamage resistance ability, were grown by Top-seeded solution growth method, and the appropriate technological conditions （mixture ratio of raw materials, thermal field et al） were adoped. By analysis, [Li]/[Nb] ratio in crystals, attributed indirectly, is about 49.7⁴9.8%, the threshold concentration and location of Mg in doped near stoichiometric LiNbO₃ crystals are confirmed, and the location model is given. The defference of Curie temperature between crystals’head and tail is less than 2℃. Crystals’photodamage resistance ability is increased to 10⁴W·cm^-2, diffraction efficiency is up to 60%, response time is under submicrosecond. All of these can indicate crystals’s component is uniform, and its ability of volume holographic storage and correlation is much higher than congruent Fe:LiNbO₃ crystals, which give a thick basis of volume holographic storage and correlation.Feature recognition vectors method is used to improve recognition accuracy under the circumstance of uneven diffraction efficiency, which must need the accurate position of correlation peaks. So, neighborhood variance increasing method based on neighborhood entropy is presented to allocate correlation peaks automatically, which can also solve the influence of background brightness. First two suitable neighborhoods are selected, then the value of variance increasing is calculated, other than entropy, to detect and locate target. It only takes less than 4 seconds to complete location, against correlation output image with size of 1024×768 pixels.Against large sample problems, an adaptive multi-scale edge extraction method is presented, which is used to improve the sharpness of correlation peaks. According the difference of Lipschitz exponent between noise and edge in wavelet transform domain, multi-scale edge correlation function is defined to check modulus extreme points, and then dual threshold, which is used to get binary image, is determined adaptively based on such criterion that within-class variance minimization. Volume holographic correlation systems with 1000 holograms from AR face database are realized. Recognition rate is 99.50% against stored image and all correct results are achieved towards 500 images not stored.Volume holographic correlation is not very effective under small sample size case. To solve this problem, a weighted two dimentional Fisherface algorithm is presented to extract features and reconstruct original information, which is used as a substitution of primary training samples and stored in crystals. First, two dimention principal component analysis is used to reduce training samples’dimension and obtain the most expressive features. Then, the most separative features is extracted by Fisher criterion, which is maximized the proportion of between-class dispersion and in-class one, and the impace of outline values is reduced by lighter weight. At last reconstructed image is rebuilt by such features. Correlation exprements based on ORL and Yale face database verify the effectivedness of the algorithm, and recognition rate is about 10% higher then that yielded by traditional volume holographic correlation.更多还原