Visual attention model VAMAI for adult image detection

doi:10.11834/jig.20110719

Views : 0 下载量: 254 CSCD: 0

PDF
Export
Share
Collection
Album

Visual attention model VAMAI for adult image detection
Vol. 16, Issue 7, Pages: 1226-1233(2011)
Published Online：18 July 2011，

Published：2011
DOI： 10.11834/jig.20110719
稿件说明：

移动端阅览

DOI：

Visual attention model VAMAI for adult image detection[J]. Journal of Image and Graphics, 2011, 16(7): 1226-1233. DOI： 10.11834/jig.20110719.

摘要

基于内容的敏感图像检测方法是过滤互联网上敏感信息的有效手段。然而，基于全局特征的检测方法误检率偏高，现有的基于BoW（bag-of-visual-words）的检测方法速度较慢。为了快速准确地检测敏感图像，本文提出基于视觉注意模型VAMAI（visual attention model for adult images）的敏感图像检测方法，包括构造面向敏感图像的视觉注意模型VAMAI、基于兴趣区域和SURF（speeded up robust features）的视觉词表算法、全局特征选择及其与BoW的融合三部分。首先，结合显著图模型、肤色分类模型和人脸检测模型，构造VAMAI，用于较准确地提取兴趣区域。然后，基于兴趣区域和SURF构建视觉词表，用于提高基于BoW的检测方法的检测速度与检测精度。最后，比较多种全局特征的性能，从中选择颜色矩，将它与BoW的支持向量机分类结果进行后融合。实验结果表明:VAMAI能够较准确地检测兴趣区域;从检测速度和检测精度两方面显著地提高了敏感图像的检测性能。

Abstract

Content-based adult image detection is an effective approach of filtering objectionable information proliferating on the Internet. However

the false positive is high for adult image detection based on global features; and the detection speed is low for adult image detection based on bag-of-visual-words (BoW) models. To detect adult images rapidly and accurately

we propose a novel method which includes three steps. First

we present VAMAI

the abbreviation of visual attention model for adult images

specialized in capturing the region of interest (ROI). VAMAI integrates saliency map model

skin-color model and face detection model to obtain ROI more precisely than ever before. Second

a codebook algorithm based on ROI and SURF (speeded up robust features) is proposed to improve the speed and the accuracy of BoW based adult image detection. Furthermore

we select color moments as the global feature after evaluating some commonly-used global features. Then we use “late fusion” to combine the results of color moments and BoW which are classified by supported vector machines. Experiments show that VAMAI is able to detect ROI precisely and our method can remarkably improve the performance of adult image detection in terms of speed and accuracy.