一种基于对抗正则化的图像特征提取方法

Image Feature Extraction Method Based on Adversarial Regularization

图像特征提取的质量直接影响图像分类精度,针对现有特征提取方法不能准确提取远离分类边界特征的问题.本文提出一种基于对抗正则化的图像特征提取方法,在分类目标函数中引入对抗性正则化项,其核心思想是将特征经过Dropout得到两个后验分布,分类器最大化后验分布得到分类边界,特征生成器最小化后验分布生成远离分类边界的特征,二者相互对抗鼓励生成更具有代表性的类别信息,同时探究不同概率分布度量方法对特征空间的影响.实验表明,经过对抗正则化后生成器网络提取的特征良好,该方法在不同数据集的分类准确率上有所提升.

The quality of feature extraction has a direct impact on the accuracy of image classification.In this paper,we propose a method based on the adversarial regularization,which introduces the adversarial regularization term into the classification objective function.The core idea is to get two posterior distributions by dropout.The classifier tries to maximize the posterior distribution to get the classification boundary,and the feature generator tries to minimize the posterior distribution to generate discriminative features.At the same time,this paper also explores the influence of different probability distribution measurement methods on feature space.The experimental results show that the features extracted by adversarial regularization are more discriminative,and the classification accuracy is improved in different data sets.