近年来，随着移动互联设备和社交网络的发展，相比与传统的文本文字交流，人与人之间倾向于采用更加形象化的图像进行信息交流。微博、微信、Facebook、推特等技术的兴起，使得每天都有数以亿计的图片上传到互联网上。面对着这些海量的图片信息，如何快速准确的找到用户所需要的信息，是各家互联网公司与企业所要面临的难题。传统的图像识别方法是采用人工的方法为图片添加各种文本文字信息来描述图像内容，但是这种方法既费时又费力。因此，如何让计算机自动的识别并标注这些图片，得到越来越多研究人员的关注。经过多年的研究与发展，基于内容的图像检索被提出，它是一种通过各种数学模型提取图像的颜色、纹理、形状和空间关系等特征来完成图像识别与匹配的技术。其中，形状特征的提取作为描述图像目标高层视觉特性的方法，更是当前研究的热点。当今社会正处于一个大数据时代，海量的图像信息分布在网络上的各个角落，传统的图像检索方法需要搜索整个互联网的图像进行匹配，这种方法需要消耗大量计算机的计算资源与内存资源。在这个追求效率的年代，数据挖掘的出现，能够让用户快速准确的得到有用的信息。通过聚类技术，让图像库中相似的图像聚成一簇，这样在检索的过程中，只需要在对应的簇中匹配待查找的图像。为了解决以上两个实际问题，本文中我们做了3个主要的工作：1. 提出了基于骨架图的图形图像检索方法。算法中，首先提出了一种图形骨架图的构建方法，它利用受限Delaunay三角剖分原则提取新的基于形状的采样点，并利用这些采样点构建形状骨架图；然后，用形状上下文方法提取骨架图的特征；最后，用动态规划完成骨架图特征间的匹配。实验结果表明，这种基于骨架图的形状描述符在保证检索精度的同时，有效的降低了描述符的算法复杂度，缩短了图像间的匹配时间。2. 提出了基于受限diffusion processes的图形图像检索方法。以往的检索方法只是比较两幅图像间的相似度距离，寻找最小的相似度值。而我们的这种检索是一种后验式的训练方法，首先将图像库中的所有图像当成一个数据流并把图像间的相似度距离转化为相似度概率；然后采用类Markov概率模型将图像间的相似度概率扩散到整个数据流；最后应用局部受限技术来消除噪声图像对整个数据流的影响。在实验结果中，这种后验式的训练方法使得图像检索的准确率得到进一步的提高。3. 提出了基于分解多目标进化算法的图形图像自动聚类方法。在该算法中，我们将多目标优化算法引入到图形图像的聚类上来。首先，将图形图像的聚类转化为一个特殊的图划分；然后，引入两个互为相反偏好的目标函数并通过优化这两个目标函数来完成图划分。由于基于分解的多目标进化算法在近几年多目标优化中有比较卓越的性能，在图形图像聚类算法中，我们采用基于分解的多目标进化算法作为算法框架来优化我们的目标函数。

In recent years, with the development of Mobile Internet Device and Social Networking Services, people prefer to communicate with each other by using image rather than by using textual words. Billions of image information are sent to the Internet every day with the application of Weibo, Macro-channel, Facebook and Twitter et al. It is difficult for the internet company to find the information of user needed quickly and accurately. In traditional technique, it is time-consuming and inefficient to add the word information on images by hand. So more and more researchers pay attention to recognize and label the image by computer automatically. The content-based image retrieval is the technology of image matching and recognition by applying Mathematical Modeling to abstract the feature of color, shape, texture and space relationship. The feature of shape become a research hotspot since it can describe the senior visual features of image target. Today is a big data time, It will consume mass of computer processor and memory resource by using traditional retrieval measures which searching the corresponding image on the whole Internet. This is really an era that requires stresses efficiency and speed. Technologies of data mining can help users find the useful information quickly and efficiently. The similar images will be parted into one cluster by applying the clustering method, so we just search the corresponding image cluster on the process of image retrieval. In order to deal with the above problems, three aspects are worked in our paper:1.A method of shape retrieval based on graph shape context is proposed. Firstly, a new shape graph structure is constructed by sampling the shape key points by the constrained Delauney triangle principle and building the skeleton graph by these key points; Then, the graph graph is described by shape context to abstract its feature; Finally, the corresponding graph nodes are matched and the similarity distance between pair of shape is calculated by dynamic programming. The experiments prove that our method has a perfect retrieval accuracy, at the same time, our method can reduce the complexity of shape descriptor and the run time of retrieval process.2. A method of shape retrieval based on the local constrained diffusion process is proposed. In the traditional retrieval technology, the shortest distance between pair of image is searched. In our method, firstly, the whole images are regarded as nodes of a graph; Then Markov Probabilistic Model is adopted to diffuse the similarity information between shapes to the whole data; Finally, local constrained method is employed to erase the negative effects of noisy image. In our experiments, this training method can be verified to enhance the performance of our retrieval results effectively.3. A method of shape automatic clustering based multi-objective optimization with decomposition approach is proposed. In this method, The muliti-objective optimization is introduced into the the area of shape clustering. Firstly, the shape clustering problem is transformed into graph-based partition; Then, two objective functions which have opposite clustered preference are applied and optimized to partition the graph. In recent years, the multi-objective optimization with decomposition (MOEA/D) have a excellent performance in optimal area. So, MOEA/D is adopted as a framework to complete the process of shape clustering in our method.