Abstract: This paper presents a blind image forensic algorithm for detecting forged images created by exemplar-based image completion technique. First, the zero-connectivity feature is used to describe the abnormal similarity of image block pairs caused by completion, and then the semi-trapezoid membership function is built up to translate the similarity characteristics of blocks into the fuzzy membership of tampered block. The tampered regions of forged image are localized by cutting the fuzzy set combining the location information of high membership blocks. The experimental results show that our algorithm can effectively distinguish natural images from inpainted forged images, and further localize the tampered regions.