基于核邻域保持判别嵌入的人脸识别

为了提取高维人脸图像中的非线性特征,提出一种新的非线性降维方法:核邻域保持判别嵌入算法(KNPDE).为了表示特征空间中类间邻域结构和不同类样本间的相似度,分别构建类内邻接矩阵和类间相似度矩阵.通过使用核技巧,KNPDE将邻域保持嵌入(NPE)和Fisher判别准则相结合,在保持特征空间中类内邻域结构的同时充分利用类间判别信息,从而具有更强的分类能力.在Yale和UMIST人脸库上的试验结果进一步表明了该算法的有效性.     

Image retrieval based on kernel trick and iterative quantization

This paper proposes a novel algorithm for solving the problem of data linear inseparable and low-accuracy in the image retrieval field. In order to get hash codes, the algorithm takes account of kernel trick and iterative quantization. First, the kernel trick is used to map the image data from low-dimension into high-dimension cleverly. In this way the data become linearly separable, and the trained hash codes are proved to be effective. Second, in the process of training the hash function, iterative quantization is used to quantize the image data to the closest hash codes. Finally, the quantitative error is minimized, and the hash codes are generated for image retrieval. Experimental results show that it certainly outperforms other compared hashing algorithms on two image benchmarks.     

基于局部敏感哈希算法的图像高维数据索引技术的研究

局部敏感哈希（LSH）算法是有效的高维数据索引方法之一,该算法成功地解决了“维数灾难”问题。分析了LSH算法中主要参数对索引性能的影响,在规模不同的图像数据集上应用了LSH算法,实验结果表明选择合适的参数时,其性能接近顺序搜索方法。     

一种稳健的非刚性医学影像配准方法

提出了一种基于局部线性嵌入(LLE)和改进的L-BFGS优化算法的非刚性配准新方法．该方法首先计算影像不同方向上的定序特征,用于补充传统互信息测度中缺失的空间结构信息;然后,运用LLE方法及其逆映射对高维定序特征进行降维和融合;进而结合影像灰度信息构造了一种基于混合熵的配准测度,有效保证了配准测度函数的光滑性和收敛性;最后,采用改进的L-BFGS优化方法搜索最优配准参数．多组仿真数据的测试结果表明,在噪声情况下,所提方法具有精度高、鲁棒性强的特点,优于现有几种方法．     

Deep consistency-preserving hashing

At present,most existing cross-modal hashing methods fail to explore the relevance and diversity of different modality data,thus leading to unsatisfactory search performance.In order to solve the above problem,a simple yet efficient deep hashing model is proposed,named deep consistency-preserving hashing for cross-modal retrieval that simultaneously exploits modality-common representation and modality-private representation through the simple end-to-end network structure,and generates compact and discriminative hash codes for multiple modalities.Compared with other deep cross-modal hashing methods,the complexity and computation of the proposed method can be neglected with significant performance improvements.Comprehensive evaluations are conducted on three cross-modal benchmark datasets which illustrate that the proposed method is superior to the state-of-the-art cross-modal hashing methods.     

实用图形水印算法研究

基于小波水印算法在不可见性和鲁棒性上能够很好地平衡,分别提出了针对矩形图像的水印扩展算法及其相似度定义,并进行了模拟运行,结果表明,新的水印扩展算法在图像降质较大的情况下依然有很好的不可见性;新的相似度定义有较强的抗降质提取图形水印的能力。     

基于随机旋转局部保持哈希的图像检索技术

针对基于局部保持投影（locality preserving projection,LPP）的哈希用于图像检索造成图像表征力不强、检索效率低下的问题,融合LPP及主成分分析（principal component analysis,PCA）技术,提出一种随机旋转局部保持哈希的图像检索算法。首先对样本进行PCA降维,对PCA变换矩阵进行随机旋转形成PCA降维矩阵,将原始样本在降维矩阵上进行投影,得到PCA降维样本。为充分利用样本间的相似性结构,对PCA降维样本进行LPP映射,并引入随机矩阵对特征向量进行偏移构造最终编码投影矩阵。再将原始样本投影到编码投影矩阵,得到最终的降维样本;最后对其进行哈希编码,得到有效的二进制编码用于图像检索。算法充分考虑样本间的全局和局部相似性结构,体现了样本间所蕴含的局部和全局信息,把随机旋转应用于PCA降维矩阵,减少了编码之间的量化误差,提高了图像特征的识别能力。分别在3个人脸数据集上进行性能测试实验,并与相关方法进行比较,得到了较好的效果。实验结果表明该方法是有效的。     

Data-oriented multi-index Hashing

The multi-index hashing (MIH) is the state-of-the-art method for indexing binary codes. However, it is based on the dataset codes uniform distribution assumption, and will lower efficiency in dealing with non-uniformly distributed codes. In this paper, we propose a data-oriented multi-index hashing method. We first compute the correlations between bits and learn adaptive projection vector for each binary substring. Then, instead of using substrings as direct indices into hash tables, we project them with corresponding projection vectors to generate new indices. With adaptive projection, the indices in each hash table are nearly uniformly distributed. Besides, we put forward an entropy based measurement to evaluate the distribution of data items in each hash table. Experiments conducted on reference large scale datasets show that compared to the MIH the time performance of our method can be 36.9％～87.4％ better .     

RA码译码简化算法的研究

RA码BP译码算法，可以实现线性时间译码，然而具有较高的复杂度。为了便于硬件实现，首先提出将最小和算法应用于RA码，然后对最小和算法进行了改进，提出了归一化算法和偏移算法，使循环译码算法在复杂度和性能之间取得了较好的折衷。仿真结果表明：归一化算法和偏移算法复杂度低，性能接近BP算法。     

知识保持的嵌入方法

考虑了一种带有数据领域知识的降维问题。这里领域知识是指关于数据的一些额外监督信息,如类别标号以及比标号弱的样本间相似性和不相似性约束等。其中,约束可以从标号中产生,但反过来从约束中却得不到标号信息,因而约束比标号更一般。另外,在图像检索等实际应用中,约束比标号更容易获取。鉴于此,本文主要研究基于约束的降维问题。提出了一种有效利用约束进行降维的约束保持嵌入算法（constraint preserving embedding, COPE）,将其纳入到图嵌入统一框架之中并指出与同类方法的关系。进一步,通过引入无标记样本提出了半监督COPE算法;提出核COPE以揭示数据中的非线性结构。最后,在人脸识别、图像检索及半监督聚类等一系列实验中的结果验证了算法的有效性。