【作者】 陈刚；

【导师】 陈晓云；

【作者基本信息】 福州大学 ， 应用数学， 2013， 硕士

【摘要】 随着信息技术的不断进步,图像作为了解世界最直观快速的一种信息媒介,在文化的传播与发展中占据了极为重要的地位。纹理图像是一种特殊的图像,其图像分布和信息传递的形式决定了它与一般图像在分析和描述上的不同。如何有效地将不同纹理进行归类,对图像的存储、管理和检索有着极大的意义。本文着重研究纹理图像的聚类问题,聚类的无监督特性为海量数据处理提供了便利。在对纹理图像特征进行聚类的问题上,单一算法虽然简单便捷,其准确率却往往不尽人意。对多个结果进行集成可以有效地弥补这一缺陷,不过已有的集成算法未能在计算复杂度和聚类准确率间达到一个有效的平衡。此外,在进行聚类之前,也必须采用具有较强概括性和较低维数的描述手段来反映图像的纹理特征。CS-LBP就具有计算复杂度小,描述性强等优点,但其在人眼视觉以及旋转方面仍存在一定的缺陷。本文就是基于上述思想来进行研究的。首先,本文将模糊理论引入到纹理谱中,设计了FCS-LBP来符合人眼识别的特性,再提出分块主纹理谱方法过滤图像的噪声。其次,考虑到已有的CS-LBP (基于中心对称的LBP)在旋转鲁棒性上的不足,提出了ECS-LBP以及移位叠加法来增强纹理谱的抗旋转能力。最后,提出了WVMC(基于最大内聚度的加权投票法)来进行聚类集成,优化了基于投票的聚类集成中的基准选择方案,并将数据点在不同聚类成员中与所划分簇中心的距离作为权值进行投票。由于该方法计算简单快速,能够应用于大规模数据集。实验表明,本文提出的两种纹理谱方法与CS-LBP相比,分别在聚类准确率和旋转鲁棒性上有了较大的提高,而WVMC聚类集成算法与传统的投票法相比,具有较高的聚类准确率和较低的时间复杂度,实用性较强。更多还原

【Abstract】 With the continuous advancement of information technology, as intuitive and fast information media for human to understand the world, image occupy a very important position in the spread and development of culture. Texture image is a kind of special image, which is different from general image in analysis and description, due to its image distribution and the form of information transmission. How to classified different textures effectively has a great significance on storage, management and retrieval of image. This paper focuses on texture images clustering. The unsupervised feature of clustering brings convenience to massive data processing.Single clustering method is simple and convenient in the clustering of texture image features, but it always has low accuracy rate. Although integration of multiple clustering results can effectively compensate for this deficiency, existing clustering ensemble algorithm failed to keep a balance between computational complexity and accuracy rate. Before clustering, it is necessary to use a description method which has strong generality and lower dimension to reflect the texture features of image. CS-LBP has the advantages of lower computational complexity and better description ability, but still has some defects in human visual system and rotation problem.Base on the above analysis, first, we introduce fuzzy theory into the texture spectrum, and the FCS-LBP is designed to match the characteristic of human eye identification. We proposed DTSP (Dominant Texture Spectrum of Partition) to filter noise. Secondly, considering the disadvantages of existing CS-LBP in rotation robustness, we design ECS-LBP and SS(shift and summation method) to enhance the anti-rotation ability of texture spectrum. Finally, we put forward WVMC to optimize benchmark options of the voting-based clustering ensemble, and vote with weights, which are generated by the distances between cluster centers and data points in different cluster members. Because this method is simple and fast, it can be applied to large-scale data sets. Simulations demonstrate that, compared with CS-LBP, our two LBP methods have great improvement in clustering accuracy and rotation robustness respectively. Compared with traditional voting methods, our WVMC clustering ensemble method has higher clustering accuracy, lower time complexity and higher practicality.更多还原