计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (14): 191-198.DOI: 10.3778/j.issn.1002-8331.1911-0438

• 图形图像处理 • 上一篇    下一篇

基于条件信息卷积生成对抗网络的图像识别

李鑫,焦斌,林蔚天   

  1. 1.上海电机学院 电气学院,上海 201306
    2.上海电机学院 继续教育学院,上海 200240
  • 出版日期:2020-07-15 发布日期:2020-07-14

Image Recognition Based on C-Info-DCGAN

LI Xin, JIAO Bin, LIN Weitian   

  1. 1.School of Electrical Engineering, Shanghai Dianji University, Shanghai 201306, China
    2.School of Continuing Education, Shanghai Dianji University, Shanghai 200240, China
  • Online:2020-07-15 Published:2020-07-14

摘要:

传统的图像识别方法需要大量有标签样本进行训练,且模型训练难以达到稳定。针对这些问题,结合条件生成网络和信息最大化生成网络的结构优势建立了条件信息卷积生成网络(C-Info-DCGAN)。模型增加图像的类别信息和潜在信息作为输入数据,然后利用Q网络去更好地发挥类别信息和潜在信息对训练的引导作用,并且利用深度卷积网络来加强对图像特征的提取能力。实验结果表明,该方法能够加快模型训练收敛速度,并有效提高图像识别的准确率。

关键词: 生成对抗网络, 信息最大化模型, 条件模型, 深度卷积网络, 图像识别

Abstract:

Traditional image recognition methods require a large number of labeled samples for training, and training model is difficult to achieve stability. Aiming at these problems, a Conditional Information Deep Convolution Generative Adversarial Network(C-Info-DCGAN) is established by combining the structural advantages of the conditional generation network and the information maximization generative adversarial network. The model adds the category information and potential information of the image as input data, and then uses the Q network to better play the guiding role of the category information and potential information on training, and uses a deep convolution network to enhance the ability to extract image features. Experimental results show that this method can accelerate the model training convergence speed and effectively improve the accuracy of image recognition.

Key words: generative adversarial network, information maximization model, conditional model, deep convolution network, image recognition