Detecting Copy-Move Forgeries via Intra-Image Relational Learning

15 concepts13 papers1 proposal

When an image is doctored by splicing an object from an external photo, it leaves behind a trail of digital breadcrumbs. Every camera sensor imparts a unique noise profile, and every lens introduces distinct artifacts. Forensic models, such as the self-supervised approach by Huh et al., exploit these discrepancies by learning a digital "fingerprint" of the imaging pipeline. If a region's fingerprint deviates from the rest of the image, it is flagged as a splice.

However, this paradigm completely breaks down when dealing with copy-move forgeries—where an object is cloned and pasted within the same image to duplicate or conceal content. As noted by Zhou et al. and Wu et al., the cloned region originates from the exact same sensor, shares the identical lighting environment, and inherits the same compression history. Because of this absolute intra-image statistical consistency, standard anomaly detection tools see the forged region as perfectly authentic. As Cui et al. emphasize, this internal uniformity makes copy-move detection a uniquely formidable challenge for classification models.

Shifting from Anomalies to Self-Similarity

To bypass the lack of external statistical anomalies, modern copy-move forgery detection (CMFD) shifts its focus from absolute anomalies to relational patterns. If the background and the forgery cannot be distinguished by their noise profiles, they must be caught by mapping identical twins within the same frame.

Early approaches tackled this by extracting robust mathematical descriptors. For instance, Feng Liu proposed combining Discrete Wavelet Transform (DWT) and Singular Value Decomposition (SVD) to capture compact texture representations that remain invariant even if the forger applies blurring or compression (An efficient algorithm for image copy-move forgery detection based on DWT and SVD; A Novel Algorithm for Image Copy-move Forgery Detection and Localization based on SVD and Projection Data). Similarly, Hu et al. utilized keypoint matching paired with an "Evolving Circular Domains Coverage" algorithm to map out matching zones.

In the deep learning era, this has evolved into dense, end-to-end relational matching. Instead of searching for exact pixel matches, networks learn to compute patch-to-patch dependencies across the entire image. For example, Islam et al. introduced a Dual-Order Attentive GAN (DOA-GAN) that explicitly models spatial locations alongside high-order patch relationships to locate duplicated regions.

To handle sophisticated geometric transformations—such as when a forger rotates or scales the cloned object before pasting it—Liu et al. and Zeng et al. designed multi-stage frameworks. By pairing deep matching with keypoint proposals or parallel feature fusion, these methods ensure that structural correspondence is captured even when the cloned target looks superficially different from its source.

The Dual-Branch Solution to Natural Textures

A naive self-similarity search introduces a massive problem: false positives. Real-world images are full of natural repetitions—think of a brick wall, a field of grass, or a row of identical windows. How does a model distinguish between a naturally recurring texture and a malicious clone?

The solution lies in dual-branch or multi-view architectures that decouple semantic meaning from fine-grained noise patterns:

  1. The Semantic Branch: Models the global context and broad image semantics using standard RGB features. This branch understands what the object is and whether its repetition is contextually plausible.
  2. The Noise/Local Branch: Extracts high-frequency noise patterns or localized details. For example, Zhou et al. fused standard RGB streams with noise features extracted via a steganalysis filter.

By looking at both views simultaneously, the network can verify if a match is merely a semantic similarity (like two different bricks on a wall) or a literal pixel-level duplicate (a cloned region). Liang et al. implemented this intuition via a dual-branch Transformer (LBRT) that refines global context with local information.

Even under weak supervision, where pixel-level masks are unavailable, Li et al. demonstrated that contrasting features across RGB and noise views allows a network to successfully isolate copy-move regions by learning multi-level relations without explicit dense labels.

Disambiguating Source from Target

Finding two identical regions is only half the battle. To provide actionable forensic proof, a system must determine which region is the authentic source and which is the manipulated target. Because both regions share the same underlying noise and texture, this requires mining incredibly subtle edge inconsistencies.

When a forger pastes a cloned region, they must blend its boundaries into the new background. This process inevitably introduces minute artifacts—such as interpolation blurring, anti-aliasing, or local contrast mismatches—along the target's perimeter.

[Source Region] ───> Perfect statistical consistency with background
[Target Region] ───> Identical internal texture, BUT exhibits boundary blending artifacts

Wu et al. pioneered this source/target localization with BusterNet, utilizing a dual-branch network where one branch detects region similarity while the other specifically hunts for manipulation artifacts along the boundaries.

More recently, Wang et al. introduced IMNet (Inconsistency Mining Network), which iteratively refines custom prototypes for both source and tampered regions. By actively mining the microscopic structural and boundary inconsistencies between the two matched regions, the network confidently breaks the statistical tie, generating clean, structurally complete masks that clearly separate the original scene from the forgery.

Key Foundational Work

An efficient algorithm for image copy-move forgery detection based on DWT and SVD (2014) — contributed 1 concept elements to this direction.
Learning Rich Features for Image Manipulation Detection (2018) — contributed 1 concept elements to this direction.
A Novel Algorithm for Image Copy-move Forgery Detection and Localization based on SVD and Projection Data (2014) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 13 papers
  2. Version 1 · 2026-10-04 · 19 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_intra_image_statistical_consistency_19011b45,
  title = {intra-image statistical consistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intra-image-statistical-consistency/rdr_19011b4569b14b249ab9891614a3453f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_intra_image_statistical_consistency_19011b45_v2,
  title = {intra-image statistical consistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/intra-image-statistical-consistency/rdr_19011b4569b14b249ab9891614a3453f/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake