查询结果:   郑欣悦,黄永辉.基于VAE和注意力机制的小样本图像分类方法[J].计算机应用与软件,2019,36(10):168 - 174.
中文标题
基于VAE和注意力机制的小样本图像分类方法
发表栏目
人工智能与识别
摘要点击数
275
英文标题
FEW-SHOT IMAGE CLASSIFICATION BASED ON VAE AND ATTENTION MECHANISM
作 者
郑欣悦 黄永辉 Zheng Xinyue Huang Yonghui
作者单位
中国科学院国家空间科学中心复杂航天系统电子信息技术重点实验室 北京 100190 中国科学院大学 北京 100049    
英文单位
Key Laboratory of Electronics and Information Technology for Space System, National Space Science Center, Chinese Academy of Science, Beijing 100190, China University of Chinese Academy of Sciences, Beijing 100049, China    
关键词
小样本学习 元学习 注意力机制 图像分类
Keywords
Few-shot learning Meta-learning Attention mechanism Image classification
基金项目
中国科学院复杂航天系统电子信息技术重点实验室自主部署基金课题(Y42613A32S)
作者资料
郑欣悦,硕士,主研领域:图像识别,神经网络,迁移学习。黄永辉,研究员。 。
文章摘要
小样本图像识别是人工智能中具有挑战性的新兴领域。传统的深度学习方法无法解决样本匮乏带来的问题,模型易出现过拟合导致训练效果不佳的情况。针对以上问题,提出结合表征学习和注意力机制的小样本学习方法。通过预训练VAE(Variational Auto-encoder)从任务中学习丰富的隐特征;对提取出的隐特征构建注意力机制,使得元学习器能快速地注意到对当前任务重要的特征;将注意力模块增强之后的特征使用分类器进行图像分类。实验表明,该算法在Mini-ImageNet和Omniglot数据集上达到72.5%和98.8%的准确率,显著优于现有元学习算法的性能。
Abstract
Few-shot learning is a challenging emerging field of artificial intelligence. Traditional deep learning models cannot solve the problem caused by the lack of samples, and the model is prone to over-fitting which leads to poor performance. In order to solve this problem, this paper proposed a few-shot learning which combined representation learning with attention mechanism. We pre-trained the VAE(variational autoencoder) to learn the rich latent features from different tasks. Then, the attention mechanism was constructed for the extracted latent features so that the meta-learner could quickly notice the key features for current learning task. Finally, the feature augmented by attention model was used to classify the image using classifier. The experimental results show that the proposed method achieves 72.5% and 98.8% accuracy on the Mini-ImageNet and Omniglot datasets respectively, which significantly surpasses the existing meta-learning algorithms.
下载PDF全文