基于类编码的判别特征学习

Discriminant feature learning based on class encoder

经典的自编码模型(BAE、SAE、DAE、CAE)都是基于输出数据对原始数据的重构,提取输入信息的低维度特征,将该特征用于图像分类不一定能够取得很好的判别效果。利用标签信息,提出了堆叠判别自编码模型(SDcAE),该模型将类编码作为隐层神经元约束加入到堆叠自编码器的训练中,使得隐层学习的特征具有更好的判别能力。同时,将类编码作为判别损失加入到Softmax分类器中,提出了类编码分类器(CEC)。由于类间样本特征误差的降低,该分类器可以取得更好的训练效果,从而提高了最终分类的正确率。实验表明,堆叠判别自编码器和类编码分类器在图像分类中是有效可行的。

Based on the output data?s reconstructing the original data,classical auto encoders(BAEA SAE、DAE、CAE)can extract low dimensional features of the input information.Applying these features in image classification may not able to guarantee a good result.In this paper,we use the label information to propose a stacked discriminant auto encoder(SDcAE),which adds the class encoderas the constraint of hidden layer neuron into the training process of the stacked auto encoder.Hence,the features learned by the hidden layer have better discrimination ability.In addition,we propose a class encoding classifier(CEC),which adds the class encoder as the discriminative loss into the Softmax classifier.Due to the de-crease of the feature error of the inter-class samples,the classifier can achieve better training results,thus improving the final classification accuracy.The experimental results show that the stacked discrimi-nant auto encoder and the class encoding classifier are effective and feasible in image classification.