Defocused Image Splicing Localization by Distinguishing Multiple Cues between Raw Naturally Blur and Artificial Blur

Splicing is one common type of forgery that maliciously changes the image contents. To make the forged image more realistic, blurring operations may be conducted to partial image regions or splicing edges to promise visual consistency. Revealing the blurring inconsistency among the whole image regions contributes to the splicing detection. However, for the defocused image already containing blur inconsistency, the existing methods cannot work well. Splicing detection and localization in defocused image is a challenging problem. In this paper, we overcome this problem by distinguishing multiple cues between raw naturally blur and artificial blur. Firstly, after the overlapped image blocks partition, three kinds of feature sets are extracted based on posterior probability map, noise histogram and derivative co-occurrence matrix. Then, an effective classifier is trained to determine the blur property of each pixel. Finally, a localization map refinement is proposed by fusing color segmentation probability map to improve the quality of the locating result. Experimental results demonstrate that the proposed method is very effective to detect splicing for the defocused images. The localization accuracy also outperforms the existing methods.