基于自适应四元数奇异值分解的图像拼接检测

Image Splicing Detection Based on Adaptive Quaternion Singular Value Decomposition

图像拼接将来源不同的图像合并成一幅图,由此引起图像中光照方向、噪声等特性出现不一致的情况。目前多数方法根据拼接图像中噪声的不一致性来检测伪造区域,但是普遍对不同大小图像块的噪声估计准确性不高,导致真阳性率较低,且当噪声差异较小时会检测失败。针对该问题,提出一种基于自适应四元数奇异值分解(QSVD)的噪声估计方法。对图像进行超像素分割,利用自适应QSVD估计超像素的噪声,结合图像亮度并利用多项式拟合建立图像噪声-亮度函数,得到各超像素到该函数曲线的最小距离测度。为提高检测精确率,利用色温估计算法提取超像素的色温特征,将距离测度与色温特征相融合作为最终的特征向量,利用FCM模糊聚类定位拼接区域。在Columbia IPDED拼接图像数据集上进行实验,结果表明,该方法在未经后处理图像集上的检测TPR值较对比方法至少提升8.21个百分点,且对高斯模糊、JPEG压缩和伽马校正表现出较好的鲁棒性。

Image splicing combines images from different sources into one image,resulting in inconsistencies in the illumination direction,noise,and other characteristics of the image.Currently,most methods detect forged areas based on the inconsistency of noise in stitched images;however,the accuracy of noise estimation for image blocks of different sizes is generally not high,resulting in a low True Positive Rate(TPR),and the detection fails when the noise difference is small.To solve this problem,a noise estimation method based on adaptive Quaternion Singular Value Decomposition(QSVD)is proposed.The image is segmented by super-pixels,and the noise of these super-pixels is estimated by adaptive QSVD.Combined with image brightness,the image noise-brightness function is established by polynomial fitting,and the minimum distance measure from each super-pixel to the function curve is obtained.To improve detection accuracy,the color temperature feature of the super-pixel is extracted using a color temperature estimation algorithm.The distance measure and color temperature feature are fused as the final feature vector.The stitching region is located by FCM fuzzy clustering.Experiments on the Columbia IPDED splicing image dataset demonstrate that the detection TPR value of this method on the unprocessed image set is at least 8.21 percentage points highter than that of the comparison method.The method is robust to Gaussian blur,JPEG compression,and Gamma correction.