基于多分类器协同学习的卷积神经网络训练算法

卷积神经网络(CNN)是一类重要的深度神经网络,然而其训练过程需要大量的已标记样本,从而限制了其实际应用。针对这一问题,分析了CNN分类器的协同学习过程,给出了基于迭代进化的分类器协同训练算法CAMC。该算法结合了CNN和多分类器协同训练的优势,首先采用不同的卷积核提取出多种样本特征以产生不同的CNN分类器；然后利用少量的已标记样本和大量的未标记样本对多个分类器进行协同训练,以持续提高分类性能。在人脸表情标准数据集上的实验结果表明,相对于传统的表情特征识别法LBP和Gabor,CAMC能够在分类过程中利用未标记样本持续实现性能提升,从而具有更高的分类准确率。

CNN Training Algorithm Based on Co-studying of Multiple Classifiers

Convolutional neural network (CNN) is a kind of important deep neural network.However,amounts of labeled samples are needed in the training process of CNN,which largely limits its applications.For the problem,the co-training process of CNN was analyzed,and a co-training algorithm CAMC based on the iterative evolution of classifiers was given.The advantages of CNN and co-training of multiple classifiers are combined in CAMC.Firstly,multiple features are drawn using different convolutional kernels to build up different CNN classifiers.Then to continually improve the classification performance,some labeled samples together with many unlabeled samples are employed to co-train the multi classifiers.The experimental results based on the standard data sets of human expression demonstrates that,compared with traditional expression recognition methods LBP and Gabor,the performance of CAMC can be continually improved,thus it has higher classification accuracy.