Journal of Frontiers of Computer Science and Technology ›› 2021, Vol. 15 ›› Issue (8): 1526-1533.DOI: 10.3778/j.issn.1673-9418.2008046

• Artificial Intelligence • Previous Articles     Next Articles

Melting Reduction Auto-Encoder

SUN Yu, WEI Benzheng, LIU Chuan, ZHANG Kuixing, CONG Jinyu   

  1. 1. College of Intelligence and Information Engineering, Shandong University of Traditional Chinese Medicine, Jinan 250355, China
    2. Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao, Shandong 266122, China
    3. Qingdao Academy of Chinese Medical Sciences, Shandong University of Traditional Chinese Medicine, Qingdao, Shandong 266122, China
  • Online:2021-08-01 Published:2021-08-02

融减自动编码器

孙宇魏本征刘川张魁星丛金玉   

  1. 1. 山东中医药大学 智能与信息工程学院,济南 250355
    2. 山东中医药大学 医学人工智能研究中心,山东 青岛 266122
    3. 山东中医药大学 青岛中医药科学院,山东 青岛 266122

Abstract:

Auto-encoder (AE) is one of the simple and widely used unsupervised feature extraction algorithms of deep learning. Existing automatic encoders for image feature extraction remain some problems such as insufficient feature extraction and excessive model parameters, etc. Aiming at above problems, MRAE (melting reduction auto-encoder) is proposed for image feature extraction in this paper. Firstly, an “ablation network structure” is proposed in the algorithm. It can realize feature enrichment through feature cross fusion in the encoder and reduce feature loss and parameters of model by optimizing the decoding structure in the decoder. Secondly, a joint reconstruction loss function is designed. It calculates the reconstruction loss between feature layers to increase the relationship between feature layers and avoid the prematurity of the model. The experimental results show that the accuracy of the feature extracted by MRAE using different classifiers, such as SVM (support vector machine), K-means, and CART (classification and regression tree), is more than 97% on lung CT image datasets. The accuracy of the feature extracted by MRAE using fully connection is more than 90% on the CvD (cats vs. dogs) dataset.

Key words: auto-encoder (AE), feature extraction, melting reduction auto-encoder (MRAE), ablation network structure, loss function of joint reconstruction

摘要:

自动编码器(AE)是深度学习领域中一种结构简单且应用广泛的无监督特征提取算法。在图像特征提取方面,现有自动编码器普遍存在特征提取不充分、模型参数量较多等问题。针对上述问题,提出了一种用于图像特征提取的融减自动编码器(MRAE)。首先,在该算法中提出“融减网络结构”,该结构在编码器中通过特征交叉传递实现了特征融合,在解码器中通过优化解码结构降低了特征损失并减少了模型参数量;其次,设计一种联合重构损失函数,该函数通过计算特征层之间的重构损失,在加强特征层之间联系的同时可有效避免模型早熟。实验结果表明:在肺部CT图像数据集上,基于融减自动编码器所提取的特征使用支持向量机(SVM)、K-means和分类回归决策树(CART)等分类器,肺炎筛查准确率均在97%以上;在CvD数据集上,基于融减自动编码器所提取的特征使用全连接分类的准确率均在90%以上。

关键词: 自动编码器(AE), 特征提取, 融减自动编码器(MRAE), 融减网络结构, 联合重构损失函数