Image Tampering Region Localization Method Based on Pseudo Label Driver and Dynamic Multi branch Decoding

In the modern world, digital images have become the primary language of visual truth. They serve as evidence in courtrooms, records in medical diagnoses, and the foundation of news reporting. Yet, the very tools that allow us to capture these moments also make it incredibly easy to alter them. A person can cut a face from one photo and paste it onto another, remove an unwanted object, or duplicate a section of a landscape to fill a gap. These manipulations are often so skillfully done that the human eye cannot spot the difference. The result is a growing crisis of trust, where seeing is no longer believing. To combat this, a field known as digital image forensics has emerged, dedicated to finding the invisible fingerprints left behind by these edits. Unlike a human observer who looks for obvious inconsistencies, forensic software scans for subtle statistical errors, noise patterns, and boundary irregularities that the camera sensor or the editing software inevitably leaves behind.

The challenge for scientists in this field is that not all edits leave the same kind of fingerprint. Some manipulations, like splicing two different photos together, create clear lines where the lighting or texture suddenly changes. Others, like copying and pasting a part of an image onto itself, are much harder to catch because the copied section comes from the same source and shares the exact same lighting and texture. Furthermore, many existing tools struggle when the edited area is tiny or when the image has been compressed or resized, which blurs the very clues the software needs to find. A new study by researchers at Jimei University and Xiamen Security Technology Vocational College proposes a solution that treats these different types of edits not as a single problem, but as a collection of distinct puzzles requiring different strategies.

The researchers developed a system that acts like a multi-tool for image analysis, designed to adapt its approach based on the specific type of tampering it encounters. Instead of forcing every image through a single, rigid process, their method first examines the image to understand what kind of manipulation might be present. It then activates different specialized pathways within its software to hunt for the specific clues associated with that type of edit. For example, if the system suspects that a piece of the image has been copied and pasted from elsewhere within the same picture, it switches to a mode that looks for matching patterns between two different areas. If it suspects a cut-and-paste job from a different photo, it focuses on the edges where the two images meet to find discontinuities in texture or lighting.

A major hurdle the team had to overcome was the lack of complete training data. In many cases, experts can mark where the fake part of an image is, but they cannot easily mark where the original, copied part came from, especially if the image was not originally labeled with that information. To solve this, the researchers created a method that allows the computer to teach itself. The system makes an educated guess about where the original source of a copied section might be, creates a temporary label for it, and then uses that label to refine its own understanding. It does this with a built-in confidence check; if the guess seems shaky or unreliable, the system gives it less weight, preventing it from learning from its own mistakes. This allows the software to learn the complex relationship between a copied region and its source without needing a human to point out every single detail beforehand.

The system also pays special attention to the edges of the manipulated areas. When a region is tampered with, the boundary between the fake and the real is often the most telling part, yet it is also the easiest to lose when an image is shrunk or compressed. The researchers added a specific layer of analysis dedicated to sharpening these boundaries, ensuring that the final map of the tampered area is not just a blurry blob, but a precise outline that matches the reality of the edit. This attention to detail is crucial for small edits, which might otherwise be swallowed up by the vast amount of untouched background in a photo.

When the researchers tested their new method against existing technologies, the results were significant. In tests where the software was trained on one set of images and then asked to find tampering in a completely different set of images, it achieved a success rate of 75.4% on splicing tasks and 52.7% on copy-move tasks. These numbers represent a clear improvement over previous methods, particularly in the difficult area of copy-move detection, where the source and target are so similar that they often fool other systems. When the software was tested on images from the same dataset it was trained on, its accuracy soared to nearly 99%, demonstrating that it can learn the specific patterns of a dataset very effectively.

The study also showed that the system is robust against real-world damage. Images in the real world are often compressed, resized, or blurred before they are examined. The new method maintained high performance even when the images were subjected to these common degradations, holding its ground where other systems began to fail. By combining a flexible, multi-pathway approach with a self-correcting learning mechanism, the researchers have created a tool that is better equipped to handle the diverse and evolving nature of digital forgery. This work suggests that the future of image forensics lies not in a single, all-purpose detector, but in adaptive systems that can recognize the unique signature of each type of manipulation and respond with the right strategy.