融合复合纹理特征的图像拼接检测

杨婧譞

摘 要：为解决传统图像拼接检测算法对图像内容、光照变化等鲁棒性不强问题，提出一種基于多种纹理特征融合的图像拼接检测方法。对二维灰度图像执行非下采样轮廓波变换（NSCT），以获得包含图像纹理特征的一系列子带图像。对在水平和垂直方向进行差分处理的低频子带图像以及4个高频图像，获取韦伯局部描述符（WLD）纹理和局部三值模式（LTP）纹理。将WLD纹理与灰度共生矩阵结合，得到像素点强度、梯度与灰度之间的关系;再将LTP纹理与灰度共生矩阵结合，得到无噪声和光照影响的像素点灰度间关系;最后分别提取WLD值共生矩阵和LTP值共生矩阵的对比度、相关性、相异度、熵、能量等5个特征，并融合成特征向量，使用RBF神经网络分类。该方法在哥伦比亚彩色图像库上检测准确率达到了95.7%。

关键词：复合纹理特征;非下采样轮廓波变换;韦伯局部描述符（WLD）;局部三值模式（LTP）;灰度共生矩阵;图像检测

DOI：10. 11907/rjdk. 182731

中图分类号：TP317.4文献标识码：A文章编号：1672-7800（2019）001-0178-05

Abstract： Aiming at the problem that traditional image splicing detection algorithm is not robust to image content and illumination change， an image splicing detection method based on multi-texture feature fusion is proposed. In this method， non-subsampling contour wave transform （NSCT） is performed on two-dimensional gray images to obtain a series of sub-band images with image texture features. Then the low frequency sub-band images were processed in the horizontal and vertical directions， and the four high frequency images were used to obtain the Weber local descriptor （WLD） texture and local three-valued model （LTP） texture. Then WLD texture and gray level co-occurrence matrix are combined to obtain the relationship among pixel intensity， gradient and grayscale; LTP texture and grayscale co-occurrence matrix are combined to obtain the relationship among the noiseless pixels affected by light. Finally the five features including WLD value symbiosis matrix， LTP co-occurrence matrix phase contrast ratio， correlation， entropy and energy are extracted and integrated into feature vectors. RBF neural network is employed for classification.

0 引言

随着计算机产业蓬勃发展以及智能手机广泛普及，图像处理的软硬件变得容易获得和使用，如Photoshop、美图秀秀等。修图成为人们的一种日常休闲娱乐，而轻而易举篡改图片也给图谋不轨的人创造了机会。近年来确实频频出现篡改或伪造图像被应用于新闻媒体、科学发现、摄影比赛和法庭物证等案例，无疑极大影响了社会稳定性。因此，在人们无法通过肉眼识别图像真实性的情况下，检测图像真实性的技术变得尤为重要。……