Generic Image Manipulation Localization Based on Boundary Artifact Search and Frequency Domain Artifact Mining

The development of multimedia technology has made image tampering and forgery easier. Tampering images carrying false information has had a huge impact on information security and media social fields. It is urgent to study the recognition and detection of tampered images. The core of the Image Manipulation Detection task is to accurately detect and locate the tampered areas in forged images. The current image tamper detection methods are often effective for detecting restrictive tampered images, but the detection effect for universal tampered images is not satisfactory. For universal image tamper detection, BFFNet utilizes two detection clues, edge features, and frequency domain characteristics, to mine frequency domain artifact traces, discover and utilize boundary artifacts hidden in edge information, and reduce the impact of semantically irrelevant objects and redundant boundaries on image tamper detection tasks. The collected frequency-domain artifacts highlight the changes caused by tampering with pixels in the image's frequency domain and further explore and utilize frequency-domain tampering artifacts to learn tampering artifact features. BFFNet unifies edge cues and frequency domain cues in an end-to-end network for joint optimization and exhibits good performance on multiple images tampering datasets.