目的 针对壁画图像具有较大类内差异以及具有较强背景噪音的特点，提出一种分组多实例学习的策略，实现对不同年代风格的壁画图像分类。方法 将样本空间划分为不同的子空间，每一个子空间中的所有训练样本训练分类器模型，测试阶段根据测试样本落到的子空间来选择不同的分类模型对测试样本进行分类。在各个子空间训练分类器时，将每一幅壁画图像样本看做多个实例的组成，采用多实例学习的方式来训练分类器。训练过程中，引入隐变量用于标识每一个实例，隐变量的存在使得分类器的优化问题不是一个凸问题，无法用梯度下降法去直接求解，采用迭代的方式训练Latent SVM作为每一个子空间的分类器。结果 实验结果表明本文方法在壁画图像的分类上与传统方法相比提高了平均5%的精度。结论 本文分组多实例学习的策略在壁画分类问题中能够较大程度地解决图像的类内差异以及背景噪音对分类结果造成的影响。

Objective Mural images have large intra-class variance and strong background noise. We present a clustered multiple instance learning strategy to classify murals of different periods and styles. Method We divide the sample space into several different sub-spaces,and a classification model is trained for each sub-space with training samples falling into this sub-space. In the testing stage,we choose a classification model for the testing sample according to the sub-space it falls into. In each classifier's training,we treat each mural image sample as a "bag" which contains a set of instances,and we use multiple instance learning to train the classifier. In the training process,we introduce hidden variables to identify each instance,the presence of hidden variables makes the classifier's optimization problem not convex which cannot be directly solved using a gradient descent. In this paper we use an iterative process to train Latent SVM(support vector machine) as the classifier for each sub-space. Result The experimental results indicate that our classification model can improve the classification accuracy of mural images by about 5% with comparison to the baseline method. Conclusion The strategy proposed in this paper can greatly reduce the impact of the intra-class variance and background noise brings to the classification result of mural images.