Computer Engineering and Applications ›› 2018, Vol. 54 ›› Issue (20): 139-144.DOI: 10.3778/j.issn.1002-8331.1707-0012

Previous Articles     Next Articles

Deep convolutional neural networks model based on connecting intermediate layers

YUAN Mingxin1,2, ZHANG Limin1,2, ZHU Youshuai1, JIANG Feng1, JIANG Yafeng2   

  1. 1.School of Mechanical Engineering, Jiangsu University of Science and Technology, Zhenjiang, Jiangsu 212003, China
    2.Zhangjiagang Industrial Technology Research Institute, Jiangsu University of Science and Technology, Zhangjiagang, Jiangsu 215600, China
  • Online:2018-10-15 Published:2018-10-19

联合中间层的深度卷积神经网络模型

袁明新1,2,张丽民1,2,朱友帅1,姜  烽1,江亚峰2   

  1. 1.江苏科技大学 机械工程学院,江苏 镇江 212003
    2.张家港江苏科技大学产业技术研究院,江苏 张家港 215600

Abstract: To solve the problems that the traditional Convolution Neural Networks(CNN) models usually use the output of the last layer as a feature representation, and fail to make full use of the intermediate layers of models, a CNN model based on connecting intermediate layers (Intermediate Layers Connected-CNN, ILC-CNN) is presented. Firstly, the front convolutional layer, the middle convolutional layer and the terminal convolutional layer of AlexNet are connected to form a more representative feature by deep connection. Then, the features which represent the whole image are obtained by pooling layer and full connected layer. In addition, the validity of the features obtained from the intermediate layers is guaranteed by training through an assistant classifier. The experimental results show that the proposed model is effective in the task of image classification and recognition, features extracted from the model are more recognizable, and the proposed model has higher recognition accuracy than other models.

Key words: image classification, image recognition, convolutional neural networks, deep connection, intermediate layers

摘要: 针对当前卷积神经网络(Convolutional Neural Networks,CNN)模型通常将网络最后一层的输出作为特征表示,未能充分利用网络中间层的不足,提出了一种联合中间层的CNN模型(Intermediate Layers Connected-CNN,ILC-CNN)。该模型以AlexNet为基础,首先联合前、中、末端卷积层,通过深度连接方式连接;接着通过池化层、全连接层等操作得到描述图像的特征向量;通过辅助分类器训练方式保证了中间层特征的有效性,使模型得以成功训练。测试结果表明,该模型在图像分类与识别任务中效果显著,其提取的特征更具辨识度,具有比其他模型更高的识别精度。

关键词: 图像分类, 图像识别, 卷积神经网络, 深度连接, 中间层