计算机应用 ›› 2017, Vol. 37 ›› Issue (9): 2705-2711.DOI: 10.11772/j.issn.1001-9081.2017.09.2705

• 应用前沿、交叉与综合 • 上一篇    下一篇

基于三通道卷积神经网络的纹身图像检测算法

许庆勇1,2, 江顺亮1, 徐少平1, 葛芸1, 唐祎玲1   

  1. 1. 南昌大学 信息工程学院, 南昌 330031;
    2. 南昌大学 经济管理学院, 南昌 330031
  • 收稿日期:2017-03-21 修回日期:2017-05-18 出版日期:2017-09-10 发布日期:2017-09-13
  • 通讯作者: 许庆勇,xyongle@ncu.edu.cn
  • 作者简介:许庆勇(1982-),男,山东成武人,讲师,博士研究生,CCF会员,主要研究方向:机器学习、图像处理;江顺亮(1965-),男,江西丰城人,教授,博士生导师,博士,主要研究方向:人工智能;徐少平(1976-),男,江西九江人,教授,博士,主要研究方向:图像处理、计算机视觉;葛芸(1983-),女,江西高安人,讲师,博士研究生,主要研究方向:数字图像处理;唐祎玲(1977-),女,浙江奉化人,讲师,博士研究生,主要研究方向:智能计算、机器学习。
  • 基金资助:
    国家自然科学基金资助项目(61662044)。

Tattoo image detection algorithm based on three-channel convolution neural network

XU Qingyong1,2, JIANG Shunliang1, XU Shaoping1, GE Yun1, TANG Yiling1   

  1. 1. School of Information Engineering, Nanchang University, Nanchang Jiangxi 330031, China;
    2. School of Economics & Management, Nanchang University, Nanchang Jiangxi 330031, China
  • Received:2017-03-21 Revised:2017-05-18 Online:2017-09-10 Published:2017-09-13
  • Supported by:
    This work is supported by the National Natural Science Foundation of China (61662044).

摘要: 针对纹身图像的特点和卷积神经网络(CNN)在全连接层对图像特征抽取能力的不足问题,提出一种三通道的卷积神经网络纹身图像检测算法,并进行了三方面的改进工作。首先,针对纹身图像的特点改进图像预处理方案;其次,设计了一个基于三通道全连接层的卷积神经网络进行特征提取,并对特征建立索引,有效地提高了网络对不同尺度下空间信息的提取能力,实现了对纹身图像的高效检测;最后,通过两个数据集验证了算法的泛化能力。实验结果表明,对NIST数据集所提预处理方案比Alex方案有总正确率提高0.17个百分点,纹身图像正确率提高0.29个百分点。在所提预处理方案下,提出的算法在标准的NIST纹身图像集上具有明显的优势,正确率从NIST公布的最优值96.3%提高到99.1%,提高了2.8个百分点;相对于传统的CNN算法,正确率从98.8%提高到99.1%,提高了0.3个百分点。在Flickr数据集上也有相应的性能提升。

关键词: 深度学习, 卷积神经网络, 纹身图像, 图像检测

Abstract: According to the characteristics of tattoo images and the insufficient ability of the Convolutional Neural Network (CNN) to extract the image features in the full connection layer, a tattoo image detection algorithm based on three-channel CNN was proposed, and three aspects of improvement work were carried out. Firstly, the image preprocessing scheme was improved for the characteristics of tattoo images. Secondly, a CNN based on three-channel fully connected layer was designed to extracted and index the features. The spatial information extraction ability of different scales was enhanced effectively, and the efficient detection of tattoo images was realized. Finally, the generalization ability of the algorithm was verified by two data sets. The experimental results on the NIST data set show that the proposed preprocessing scheme has a 0.17 percentage points increase of total correct rate and a 0.29 percentage points increase of correct rate for tattoo images than Alex scheme. Under the proposed preprocessing scheme, the proposed algorithm has obvious advantages on the standard NIST tattoo image set. The correct rate of the proposed algorithm reaches 99.1%, which is higher than 96.3%, the optimal value published by NIST; and 98.8%, obtained by traditional CNN algorithm. There is also a performance improvement on the Flickr data set.

Key words: deep learning, Convolutional Neural Network (CNN), tattoo images, image detection

中图分类号: