Biorthogonal wavelet transform‐based adaptive segmentation and statistical keypoint matching approach for image forgery detection

Abstract Detecting copy–move forgeries in digital images, especially those with smooth, small, and geometrically distorted regions, remains challenging. Existing methods often cannot efficiently locate forged regions in such cases. Hence, the adaptive kernel function‐based superpixel segmentation (AKSS) algorithm based on the biorthogonal wavelet transform is proposed in this study. This method increases forgery detection efficiency and is robust against different attacks. The method has three main stages: First, the input image is segmented using the proposed AKSS algorithm. Next, features are extracted and a proposed dual‐stage keypoint matching step is performed on the segmented image. Finally, a method is proposed for forgery localization based on the matched keypoints and segments. Experiments were performed on several copy–move forgery image datasets. The proposed method obtained an average ‐score of 100% at image level and 97.45% at pixel level on three standard datasets, and an average improvement in ‐score of 1.21% was obtained when compared with existing methods. The proposed forgery detection method was also tested on online social media‐transmitted images at pixel level on four datasets and achieved an ‐score of around 95%. The results show that the proposed method outperforms existing methods under various attack scenarios.