计算机工程与应用 ›› 2016, Vol. 52 ›› Issue (21): 180-187.

• 模式识别与人工智能 • 上一篇    下一篇

基于卷积词袋网络的视觉识别

薛昆南,薛月菊,毛  亮,刘洪山   

  1. 华南农业大学 电子工程学院,广州 510642
  • 出版日期:2016-11-01 发布日期:2016-11-17

Bag of convolutional words networks for visual recognition

XUE Kunnan, XUE Yueju, MAO Liang, LIU Hongshan   

  1. School of Electronic Engineering, South China Agricultural University, Guangzhou 510642, China
  • Online:2016-11-01 Published:2016-11-17

摘要: 近年来,卷积神经网络(CNN)凭借其强大的特征学习能力在视觉识别领域取得重要进展。针对CNN全连接层对图像平移、旋转、缩放等变换比较敏感的问题,提出了一种混合模型——卷积词袋网络(BoCW-Net)。它将BoW模型嵌入CNN结构中并代替全连接层,通过端到端的方式学习特征、字典和分类器。为实现BoCW-Net整个网络的有监督学习,提出基于方向相似度的BoCW编码。同时,为充分利用中层特征和高层特征的鉴别性,将中层辅助分类器与高层分类器集成,形成主-辅集成分类器。实验结果表明:相比全连接层,BoCW表示对各种变换具有更强的不变性;主-辅集成分类器能有效融合中层、高层特征,提高BoCW-Net的识别性能;相比新近发展的CNN模型,BoCW-Net在CIFAR-10、CIFAR-100和MNIST数据库上均取得了改进的识别性能,最终分别获得4.88%、22.48%和0.21%的测试错误率。

关键词: 卷积神经网络, 卷积词袋(BoCW)表示, 主-辅集成分类器

Abstract: In recent years, Convolutional Neural Networks(CNN) have made a progress in visual recognition tasks with its powerful feature learning ability. A hybrid model called BoCW-Net is proposed to solve the problem that full-connection layer in CNN is more sensitive to image’s transformations such as translation, rotation and scale, et al. It embeds BoW model into CNN architectures and replaces the full-connection layer, while it can learn feature, dictionary and classifier in the end-to-end way. In order to realize supervised learning of whole BoCW-Net, BoCW encoding based on direction similarity is proposed. In the meanwhile, to take full advantage of the discrimination of both mid-level and high-level features, middle-level auxiliary classifier is integrated to high-level classifier to form the main-auxiliary ensemble classifier. Experimental results show that BoW model imbedded into CNN has better invariance for a variety of transformations compared with the full-connection layer. Main-auxiliary ensemble classifier can effectively fusion mid-level and high-level features to improve the recognition performance of BoCW-Net. Compared with the newly developed CNN models, BoCW-Net acquires improved recognition performance on CIFAR-10、CIFAR-100 and MNIST dataset with 4.88%, 22.48% and 0.21% final test error rate, respectively.

Key words: convolutional neural networks, Bag of Convolutional Words(BoCW) representation, main-auxiliary ensemble classifier