Local Geometry Analysis For Image Tampering Detection

In this paper, we propose a compliant scheme of image hashing that is based on gradient measurement. The proposed approach relies on a local geometry variation analysis in multi-channel images. Derivative filters are used to estimate the image gradient and a suitable representation of geometric color features is introduced. A gradient norm is provided as an intermediate hash, so we compute subsequently the magnitude of the Fourier spectrum in order to provide a final hash. The aim is to preserve accurately the structural information along with smoothing operations. We conduct experiments on the CASIA-V2 database for the tampering detection task, and the UCID database to demonstrate the hash robustness. We use the metrics of true positive rate (TPR) and false positive rate (FPR) to investigate the performance of the proposed method. A comparison is carried out using two corresponding schemes; the first operates on the quaternion discrete Fourier transform (QDFT) to take into account the image color planes, and the second exploits this transform into the log-polar domain. According to the TPR results, our method is quite robust against different content-preserving operations applied on the UCID database with regard to a predefined threshold. The FPR results over CASIA-V2 further demonstrate a superior capability of the proposed approach in detecting image forgeries.