针对当前大多数无监督图像分类方法不能对每个图像类进行特征选择和自动确定图像类别的数量问题,提出一种基于Adaboost和随机图划分的无监督图像分类方法。该方法包括两个部分:1)将图像分类问题看做是一个自动的随机图划分问题,其中图的每一个顶点代表一幅图像,通过划分形成的子图代表了图像类。再采用Adaboost算法对每一个形成的图像类进行特征选择,从而得到每类图像的表达模型。2)采用一种基于蒙特卡洛马尔可夫链(MCMC)的随机采样算法(SWC)来对图进行划分。相比传统的随机采样算法,SWC具有更快的收敛速度。在两个图像数据集上的实验结果表明,本文方法的分类性能明显优于其他现有的无监督分类法。

In this paper,we present a general framework to discover image categories automatically.The algorithm includes two parts:1)we pose the problem of category discovery as an automated graph partition task. Each graph vertex indicates an image,and a partitioned sub-graph consisting of connected graph vertices representing a clustered category. The model of each image category can be learned by stepwise feature selection using the Adaboost algorithm. 2)A MCMC-based stochastic algorithm,the Swendsen-Wang Cuts (SWC),is adopted to solve the graph partition fast. Compared to traditional random cluster sampling techniques,SWC converges faster. We apply our method on two image datasets,and the experimental results demonstrate superior performance of our method over other popular state-of-the-arts methods,including Kmeans,pLSA,and Affinity Propagation.