摘要
针对矢量量化编码的量化误差严重,而稀疏编码只是一种浅层学习模型,容易导致视觉词典对图像特征缺乏选择性的问题,提出了一种基于深度学习特征编码模型的图像分类方法。首先,采用深度学习网络无监督的受限玻尔兹曼机(RBM)代替传统的K-Means聚类及稀疏编码等方法对SIFT特征库进行编码学习,生成视觉词典;其次,对RBM编码添加正则化项分解组合每个特征的稀疏表示,使得生成的视觉单词兼具稀疏性和选择性;然后,利用训练数据的类别标签信息有监督地自上而下对得到的初始视觉词典进行微调,得到图像深度学习表示向量,以此训练SVM分类器并完成图像分类。实验结果表明,本文方法能有效克服传统矢量量化编码及稀疏编码等方法的缺点,有效地提升图像分类性能。
For the serious quantization error in vector quantitation coding, the sparse coding is only a shallow learning model which caused the codeword lack selectivity for image features. In this paper, an image classification method based on deep learning coding model was proposed. Firstly, the deep learning network unsupervised RBM was used to encode SIFT features and generate visual diction- ary instead of the traditional K-means elustering. Then, the unsupervised RBM learning was steered by using a regularization scheme, which decomposes into a combined prior for the sparsity of each feature' s representation as well as the selectivity for each eodeword. Finally, the initial dictionary was fine-tuned to be discriminative through the supervised learning from top-down labels. To train SVM classifter and complete image classification,the representation features based on image deep learning were obtained. The experimental resuits demonstrated that the proposed method can overcome the disadvantage of vector quantization coding and sparse coding. Moreover, the classification performance can be boosted effectively.
出处
《工程科学与技术》
EI
CAS
CSCD
北大核心
2017年第1期213-220,共8页
Advanced Engineering Sciences
基金
国家自然科学基金资助项目(61379152
61301232)
全军军事学研究生课题资助项目(YJS1062)
关键词
图像分类
视觉词典模型
深度学习
稀疏编码
受限玻尔兹曼机
image classification
bag of visual words model
deep learning coding model
sparse coding
restricted Bohzmann machine