Abstract: The Semi-Supervised Learning（SSL） method of Laplacian regularization opens a bright avenue by studying the classification problem of a large number of images. However, Laplacian method biases the classi?cation function toward a constant function and possibly results in poor generalization. So the Hessian regularized Logistic Regression（HesLR） method for image classification is proposed. Hessian regularization can well predict the data points beyond the boundary of the domain. Extensive experiments on the MIR Flickr dataset validate the effectiveness of the proposed method by comparing it with baseline algorithms, including SVM, the Logistic regression method and the Laplacian regularized Logistic regression method.

Key words: Semi-Supervised Learning（SSL）, Hessian, kernel Logistic regression, image classification

摘要： 针对大数据量的图像分类问题，Laplacian正则化的半监督学习方法获得了广阔的应用前景。然而Laplacian正则化使分类函数趋向于常数函数而易导致较差的推测能力。提出了基于Hessian正则化的Logistic回归模型用于图像分类，Hessian正则化可以较好地预测区域之外的数据点。在MIR Flickr数据库上进行图像分类实验，与SVM、Logistic回归和Laplacian正则化的Logistic回归方法相比，Hessian正则化的Logistic回归模型更有效。

关键词: 半监督学习, Hessian, Logistic核回归, 图像分类