结合主动反馈的图像多分类框架

Efficient active feedback scheme of image multi-class classification

为了解决图像语义分类中的训练数据不对称、小样本训练和噪声数据这3个难题,提出结合主动反馈的图像多分类框架。该框架将主动选择的策略应用到图像的多分类中,通过主动的选择出不确定的图片给用户手动标记,扩大训练图片集,提高分类的精度。为了验证该框架的有效性,提出一种有效的结合主动选择的图像多分类算法,即结合投票的DDAGSVM(decision directed acyclic graph support vector machine)算法。该算法提出了新的主动选择策略,即结合投票和旁移机制的主动选择策略。实验结果表明,该算法能有效应用到图像多分类中,比DDAGSVM和采用普通主动选择策略的DDAGSVM具有更高的分类的精度。

In order to solve three difficulties of image classification,including asymmetry of training data,small sample issue and noise sample problem,efficient active feedback scheme of image multi-class classification which introduce active selecting technique into image multi-class classification is proposed.By actively selecting doubtful images for users to label,more training samples can be gotten and more accurate classification can be done.In order to validate the scheme,an image multi-class classification algorithm combining with active selecting is fulfilled,which is named as DDAG SVM（decision directed acyclic graph support vector machine） with voting.Experiments show that the algorithm has more accuracy than DDAG SVM and DDAG SVM with normal selecting strategy,so it is effi-cient and the proposed scheme is also good for image multi-class classification.