Image Splicing Detection Based on Adaptive Quaternion Singular Value Decomposition

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.