计算机科学与探索 ›› 2020, Vol. 14 ›› Issue (2): 252-259.DOI: 10.3778/j.issn.1673-9418.1909079

• 网络与信息安全 • 上一篇    下一篇

多数据集深度学习模型的修图处理识别

杨滨,陈先意   

  1. 1. 江南大学 设计学院,江苏 无锡 214122
    2. 江南大学 轻工过程先进控制教育部重点实验室,江苏 无锡 214122
    3. 南京信息工程大学 计算机与软件学院,南京 210044
  • 出版日期:2020-02-01 发布日期:2020-02-16

Image Modification Recognition Based on Multi-Dataset Deep Learning Model

YANG Bin, CHEN Xianyi   

  1. 1. School of Design, Jiangnan University, Wuxi, Jiangsu 214122, China
    2. Key Laboratory of Advanced Process Control of Light Industry of Ministry of Education, Jiangnan University, Wuxi, Jiangsu 214122, China
    3. School of Computer and Software, Nanjing University of Information Science and Technology, Nanjing 210044, China
  • Online:2020-02-01 Published:2020-02-16

摘要:

图像处理软件的飞速发展,带动了移动应用领域一大批修图、美化应用的兴起。但是修图、美化软件的快速发展和普及也带来了一些社会问题和安全问题,如网恋对象严重失真,摄影作品造假等。针对手机中的修图处理APP软件,提出一种基于多数据集特征学习的神经网络模型,并给出其网络拓扑结构。区别于传统的多个神经网络并行操作,提出的网络模型具有共享模型参数的特征,能同时对多个特征数据集进行深度学习,使检测程序具备多特征识别能力。此外,还提出了一种针对多任务网络模型的损失函数,以增强深度特征学习的能力。实验结果表明,提出方法的准确率较传统方法有较大提升,同时泛化性能优越,能识别出经过多种美图、修图软件修复过的图像。

关键词: 多数据集学习, 修图识别, 深度学习, 神经网络设计, 图像处理

Abstract:

The rapid development of image processing software has led to the rise of a large number of repair and beautification applications in the field of mobile applications. However, their popularity has also brought some social problems and security problems, such as the serious distortion of online lovers, the forgery of photographic works, etc. For mobile phone image modification APP software, a neural network model based on multi-dataset learning is proposed and its network topology is given. Different from the traditional parallel operation of multiple neural networks, the proposed network model can share the model parameters, and learn multiple data sets at the same time. Thus, the detection application has the ability of multi-feature recognition. In addition, a loss function for multi-task network model is proposed to enhance the ability of deep feature learning. The experimental results show that the accuracy of the proposed method is higher than that of the traditional method. The generalization performance of the proposed method is superior, and it can recognize the image which has been repaired by many kinds of software.

Key words: multi-dataset learning, image modification recognition, deep learning, design of neural network, image processing