图书情报工作 ›› 2019, Vol. 63 ›› Issue (11): 96-107.DOI: 10.13266/j.issn.0252-3116.2019.11.011

• 知识组织 • 上一篇    下一篇

基于深度学习CNN模型的图像情感特征抽取研究

李志义1, 许洪凯1, 段斌2   

  1. 1. 华南师范大学经济与管理学院 广州 510006;
    2. 华南师范大学信息光电子科技学院 广州 510006
  • 收稿日期:2018-08-08 修回日期:2018-12-19 出版日期:2019-06-05 发布日期:2019-06-05
  • 作者简介:李志义(ORCID:0000-0001-6407-2554),副教授,硕士,硕士生导师,E-mail:Leeds@scnu.edu.cn;许洪凯(ORCID:0000-0002-3304-4312),硕士研究生;段斌,本科生。
  • 基金资助:
    本文系国家社会科学基金项目"基于表示学习的跨模态检索模型与特征抽取研究"(编号:17BTQ062)研究成果之一。

Research on Image Emotion Feature Extraction Based on Deep Learning CNN Model

Li Zhiyi1, Xu Hongkai1, Duan Bin2   

  1. 1. Economic & Management College of South China Normal University, Guangzhou 510006;
    2. Information & Photoelectric Science College of South China Normal University, Guangzhou 510006
  • Received:2018-08-08 Revised:2018-12-19 Online:2019-06-05 Published:2019-06-05

摘要: [目的/意义]以用户情感为线索的图像检索已成为机器学习研究的热点,但图像情感特征标注的语料数据多来源于对图像低层特征的抽取,从而导致图像检索过程单一化和程式化。本文提出了一种基于深度学习的图像情感特征抽取的算法,将图像底层特征融合到图像的高层情感语义当中,为实现图像的情感语义检索提供了参考。[方法/过程]利用改进的卷积网络模型,将数据集图像的颜色、纹理作为输入,经多层运算自动提取图像的情感信息,并通过反向传播算法计算出改进后模型的情感检索准确率,构造出准确率较高且过拟合程度低的图像情感特征提取模型。[结果/结论]应用改进的卷积神经网络模型,实现了对图像情感特征的抽取,相较于原模型提升了10%的检索准确率。

关键词: 深度学习, 图像, 情感特征, 抽取, 卷积神经网络

Abstract: [Purpose/significance] Image retrieval based on user emotion has become a hotspot in machine learning research. However, the corpus data of image sentiment feature annotation is mostly derived from the extraction of low-level features of images, which leads to the simplification and stylization of image retrieval process. The algorithm of image emotion feature extraction based on deep learning fuses the underlying features of the image into the high-level emotion semantics of the image, which provides a reference for the emotional semantic retrieval of images.[Method/process] Using the improved convolutional network model, the color and texture of the dataset image were taken as input, the emotion information of the image was automatically extracted by multi-layer operation, and the sentiment retrieval accuracy of the improved model was calculated though the back propagation algorithm, and an image sentiment feature extraction model with high rate and low degree of over-fitting was constructed.[Result/conclusion] This paper completes the extraction of emotional features of the image through an improved deep convolutional network model, and improves the retrieval accuracy by 10%.

Key words: deep learning, image, emotional features, extraction, convolutional neural network

中图分类号: