度量学习改进语义自编码零样本分类算法

Improving Semantic Autoencoder Zero-Shot Classification Algorithm by Metric Learning

为改善零样本图像分类中相似度度量方法的鲁棒性,引入了一种用于零样本分类的度量学习方法.该方法由自编码构成,能在特征对齐后的语义嵌入空间中学习到最优的度量函数,用于计算测试样本特征和类标签的语义特征的相似度;然后利用近邻思想预测类别标签,进而避免产生不合适距离函数导致的分类错误.实验结果表明,与传统距离度量的算法相比,所提出的方法降低了识别错误率,在公开数据集AWA、CUB和Im Net-2上的分类准确率分别达到94.7%、63.7%和28.59%;同时表明了语义—视觉的映射方向比相反方向的识别准确率高出2.5%~10.1%.

To improve the robustness of similarity metric method in zero-shot learning,a new metric learning for zero-shot image classification was introduced. It is composed of autoencoders,which can learn the optimal metric function in the feature-aligned semantic embedding space. The similarity between test sample features and the semantic features of the class labels can be calculated by metric function,predicting the label of the class by the neighboring method. Thus,the classification error caused by inappropriate distance function is prevented. Compared with the traditional distance metric algorithm,the experiments show that the proposed method reduces the recognition error rate; the recognition accuracy is improved to 94. 7%,63. 7% and 28. 59% on the AWA,CUB and Im Net-2 datasets. At the same time,it was confirmed that the recognition accuracy of the semantic-visual mapping direction was 2. 5% ~10. 1% higher than the opposite direction.