计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (20): 199-205.DOI: 10.3778/j.issn.1002-8331.1909-0188

• 图形图像处理 • 上一篇    下一篇

结合模型集成与特征融合的图像拷贝检测

武光华,张旭东,葛维,孙鸽,毛财胜   

  1. 1.国网河北省电力有限公司 电力科学研究院,石家庄 050021
    2.国网河北省电力有限公司,石家庄 050000
    3.大连理工大学 软件学院,辽宁 大连 116024
  • 出版日期:2020-10-15 发布日期:2020-10-13

Image Copy Detection Based on Model Ensemble and Feature Fusion

WU Guanghua, ZHANG Xudong, GE Wei, SUN Ge, MAO Caisheng   

  1. 1.State Grid Hebei Electric Power Research Institute, Shijiazhuang 050021, China
    2.State Grid Hebei Electric Power, Shijiazhuang 050000, China
    3.School of Software Technology, Dalian University of Technology, Dalian, Liaoning 116024, China
  • Online:2020-10-15 Published:2020-10-13

摘要:

基于内容的图像拷贝检测关键在于提取的图像特征能够针对不同形式的图像拷贝攻击具有不变性。现实中拷贝攻击手段变化多样,且存在很多相似图像的干扰,目前并没有任何一种图像特征可以对抗所有不同形式的图像攻击。现有方法虽然在图像特征表示上做了很多改进,但都局限于单个特征表示。因此从特征融合的角度对提取特征进行增强,基于卷积神经网络融合图像高层特征以及低层特征以实现特征多样性,集成ImageNet预训练分类模型以及提出的距离度量模型以实现特征互补性。度量模型针对该类问题在预训练模型的基础上通过学习合适的距离度量来对抗由于图像编辑引起的特征差异,拉近拷贝图像与原始图像在特征空间的距离。实验结果表明,结合模型集成和多层深度特征融合的方式可以有效增强特征的鲁棒性,相比单一特征的检测效果提升十分明显。

关键词: 拷贝检测, 特征融合, 卷积神经网络, 度量学习, 图像检索

Abstract:

The key of content-based image copy detection is that the extracted image features are invariant to different image copy attacks. However, in real-world scenarios, copy attacks are in various forms, and there are many interferences from similar images. No image feature can resist all forms of copy attacks. Although the existing methods have made many improvements in image feature representations, they are limited to a single feature representation. Therefore, this article boosts the extracted features from the perspective of feature fusion. Firstly, based on the convolutional neural network, the high-level features and low-level features of the image are fused to achieve feature diversity. Secondly, the ImageNet pre-trained classification model and the proposed distance metric model are integrated to achieve feature complementarity. The metric model learns a suitable distance metric on the basis of the pre-trained model to resist feature differences caused by image editing, and narrows the distance between the copied image and the original image in the feature space. The experimental results show that the proposed method based on model ensemble and multi-layer feature fusion can effectively enhance the robustness of the feature, and significantly improves the detection results compared to a single feature.

Key words: copy detection, feature fusion, convolutional neural network, metric learning, image retrieval