Forensics Rectification Module

• Forensics Rectification Module is a core component that revalidates semantic proposal embeddings using detailed, multi-scale forensic evidence.
• It integrates outputs from forensic filters (SRM, Bayar, Sobel, Noiseprint++) with proposal-aware gating to adaptively refine manipulation localization.
• Empirical results demonstrate significant improvements in detection and localization metrics by combining technical evidence with contextual MLLM analysis.

The Forensics Rectification Module is a component for image manipulation localization that appears in the two-stage Propose-Rectify framework introduced in "Propose and Rectify: A Forensics-Driven MLLM Framework for Image Manipulation Localization" (Zhang et al., 25 Aug 2025). In that framework, a forensic-adapted MLLM first generates initial manipulation analysis and preliminary localization of suspicious regions, and the rectification stage then systematically validates and refines these initial proposals through multi-scale forensic feature analysis, integrating technical evidence from several specialized filters. The module is designed to bridge semantic reasoning and low-level forensic analysis: it does not treat the MLLM proposal as final, but instead re-evaluates it using explicit forensic evidence, adaptive proposal-aware gating, and progressive embedding updates (Zhang et al., 25 Aug 2025).

1. Conceptual definition and scope

Within the Propose-Rectify framework, the Forensics Rectification Module (FRM) is the technical center of the Rectify stage. Its direct input consists of initial proposal embeddings from a forensic-adapted MLLM together with comprehensive multi-feature forensic evidence from a low-level artifact extractor. Its direct output is a set of rectified embeddings that support both image-level detection and downstream segmentation (Zhang et al., 25 Aug 2025).

The module was introduced to address a specific failure mode in MLLM-based forgery localization. The underlying paper states that recent MLLM systems can leverage semantic understanding and world knowledge, but their encoders are primarily tuned for semantic understanding and are therefore insensitive to subtle, low-level forensic artifacts, including noise or compression traces. The paper also states that such systems often over-rely on textual region descriptions, can suffer from hallucinations and error propagation, and lack mechanisms for rigorous technical validation of their hypotheses (Zhang et al., 25 Aug 2025).

In this formulation, “rectification” does not denote geometric dewarping. It denotes proposal validation and refinement: the initial semantic hypothesis is re-grounded in technical evidence before classification and localization are finalized. This makes the term distinct from document dewarping or fingerprint distortion correction, even though those literatures also use “rectification” for deformation recovery (Zhang et al., 2022, Guan et al., 2024).

2. Motivation and the limitations it addresses

The FRM is motivated by a split between two established traditions in forensic vision. Traditional image forensic methods rely on handcrafted feature extractors or deep models that analyze low-level statistical inconsistencies; according to the paper, these methods can be effective in controlled settings but often lack generalization to diverse manipulation types, can suffer under noise or compression, and cannot utilize semantic context or world knowledge. MLLM-based systems, by contrast, can propose suspicious regions from context and semantics, yet may miss imperceptible forensic traces and may localize poorly when high-level visual clues are weak (Zhang et al., 25 Aug 2025).

The paper frames FRM as the mechanism that reconciles those two traditions. The MLLM contributes contextual reasoning and preliminary localization. The FRM contributes a rigorous technical validation layer based on multi-scale forensic evidence. The stated objective is to combine semantic reasoning and forensic analysis for robust, generalizable, and precise detection/localization (Zhang et al., 25 Aug 2025).

This design places the module in a broader family of forensic “rectification” strategies. In ReLoc, for example, a restoration module is optimized not only for visual quality but also with a forensics-oriented objective function, so that distorted tampering traces can be re-enhanced before localization (Zhuang et al., 2022). In ForgeryVCR, a forensic toolbox converts imperceptible traces into explicit visual intermediates for visual-centric reasoning (Wang et al., 15 Feb 2026). In ForensicsTok, the model compensates for the lack of forensic priors in standard MLLMs through Hierarchical Expert Fusion and direct tokenized mask generation (Xu et al., 23 Jun 2026). These neighboring formulations suggest that “rectification” in forensic analysis increasingly refers to modules that reintroduce technical evidence into pipelines otherwise dominated by semantic reasoning.

3. Architecture and information flow

The FRM is organized around four operations: multi-feature forensic extraction, a semantic-forensic bridge, analysis-informed feature gating, and progressive multi-scale rectification. The following summary stays close to the formulation reported in the paper.

Component	Operation	Role
Multi-Features Extractor	SRM, Bayar Filter, Sobel Filter, Noiseprint++	Builds unified forensic evidence
Semantic-forensic bridge	Multi-Head Cross-Attention	Aligns proposal embeddings with forensic features
Feature gating	Proposal-conditioned weights $[w^1, w^2, w^3]$	Reweights evidence by analysis scale
Multi-scale rectification	Convolution + MSA + MCA	Refines analysis and segmentation embeddings

The Multi-Features Extractor concatenates the outputs of several complementary forensic filters: SRM, Bayar Filter, Sobel Filter, and Noiseprint++. The paper defines the unified feature tensor as

$$F = \text{Concat}([f_{\text{SRM}(I)}, f_{\text{Bayar}(I)}, f_{\text{Sobel}(I)}, f_{\text{Np++}(I)}]).$$

This tensor $F \in \mathbb{R}^{H \times W \times K}$ is described as summarizing technical evidence from different forensic perspectives (Zhang et al., 25 Aug 2025).

The semantic-forensic bridge then aligns MLLM proposal embeddings with the patch-embedded forensic tensor. The initial proposal embeddings are extracted as

$$[e^0_{ANL}, e^0_{SEG}] = \mathcal{H}(\text{MLLM}(I, T), [pos_{ANL}, pos_{SEG}]).$$

Cross-attention between $e^0_{ANL}$ and $\text{PE}(F)$ produces proposal-aware scale weights:

$$[w^1, w^2, w^3] = \Phi(\text{MCA}(e^0_{ANL}, \text{PE}(F))).$$

These weights define the analysis-informed feature gating step:

$$F^{k} = \sigma(w^{k}) \odot F, \quad k \in \{1,2,3\}.$$

The paper emphasizes that this mechanism dynamically adapts feature emphasis depending on the MLLM’s analysis, so that the most relevant forensic aspects are prioritized for the current image and context (Zhang et al., 25 Aug 2025).

Finally, the module performs progressive rectification across three analysis scales. For each scale,

$$\hat{F}^{k} = \text{MSA}(\text{PE}(\text{Conv}_{s_k}(F^{k})))$$

and the proposal embeddings are updated by cross-attention:

$$[e_{ANL}^{k}, e_{SEG}^{k}] = \text{MCA}([e_{ANL}^{k-1}, e_{SEG}^{k-1}], \hat{F}^{k}).$$

After the last scale,

$$F = \text{Concat}([f_{\text{SRM}(I)}, f_{\text{Bayar}(I)}, f_{\text{Sobel}(I)}, f_{\text{Np++}(I)}]).$$0

The paper summarizes this as a progressive, multi-stage validation process in which the final decision is grounded in both high-level semantic reasoning and scale-adaptive forensic evidence (Zhang et al., 25 Aug 2025).

4. Multi-scale analysis and proposal-conditioned rectification

A defining property of the FRM is its explicit multi-scale forensic feature analysis. The paper states that manipulation artifacts manifest differently at different spatial extents, and the module therefore deploys three analysis scales:

• Local Scale: fine-grained analysis with 3×3 kernels
• Medium Scale: regional analysis with 7×7 kernels
• Global Scale: extended-context analysis with 9×9 kernels with dilations (Zhang et al., 25 Aug 2025)

The local scale is reported to target pixel-level inconsistencies and small artifacts. The medium scale captures regional distortions. The global scale identifies extended or contextual anomalies. In the paper’s formulation, this makes the module capable of addressing localized splices, subtle retouching, and wide-area inpainting within a single rectification pipeline (Zhang et al., 25 Aug 2025).

The scale analysis is not static. It is driven by the initial semantic proposal through the analysis-informed feature gating mechanism. Because the gating weights are computed from the MLLM analysis embedding and forensic features, the evidence used at each scale is proposal-conditioned rather than fixed. The paper characterizes this as adaptive, proposal-aware feature gating (Zhang et al., 25 Aug 2025).

This proposal-conditioned design is important conceptually. A semantic-only proposal may identify the approximate region of interest but may remain vulnerable to hallucinations or imprecise boundaries. The FRM re-examines that proposal using gated forensic evidence at local, regional, and global scales, and the updated embeddings are then passed to the classification head and to the segmentation path. The paper states that this process removes hallucinations, corrects localization errors, and grounds decisions in technical evidence (Zhang et al., 25 Aug 2025).

A plausible implication is that the FRM can be understood as an evidence arbitration layer between semantic plausibility and statistical trace analysis. That interpretation is consistent with evidence-gated reasoning in other forensic systems, although the exact implementation differs. For example, EG-SIEM and EG-SIEM-Enron confirm alerts only when multiple evidence channels align, including behavioral anomaly, communication forensics, peer-group anomaly confirmation, and Theory-of-Mind consistency (Kausar et al., 6 Jan 2026). The FRM performs an analogous arbitration in visual forgery localization, but through attention, scale-specific convolution, and embedding refinement rather than symbolic gate logic.

5. Integration with the Propose-Rectify framework

The paper presents the FRM as inseparable from the full Propose-Rectify architecture. The operational sequence is:

1. The forensic-adapted MLLM analyzes the image semantically and generates preliminary tampered-region proposals.
2. The FRM receives those proposal embeddings together with the unified forensic feature tensor.
3. The FRM performs multi-scale, context-adaptive forensic validation and proposal refinement.
4. The rectified embeddings then guide the enhanced segmentation path for mask generation (Zhang et al., 25 Aug 2025).

The paper explicitly states that the initial proposals are not considered final. This design choice distinguishes the framework from pipelines in which an MLLM’s textual or semantic region description is passed downstream with little or no correction. Here, the semantic stage supplies a hypothesis, and the FRM tests that hypothesis against technical evidence before the final decision is formed (Zhang et al., 25 Aug 2025).

This placement is consistent with a broader pattern in recent work on forensic reasoning with MLLMs. ForgeryVCR replaces text-centric Chain-of-Thought with Visual-Centric Reasoning, using ELA, Noise Print++, FFT, and Zoom-In to materialize imperceptible traces into explicit visual intermediates (Wang et al., 15 Feb 2026). ForensicsTok removes exogenous segmentation bottlenecks by reformulating localization as autoregressive token generation and injecting expert forensic features through Hierarchical Expert Fusion (Xu et al., 23 Jun 2026). In all three cases, low-level forensic evidence is reintroduced after or alongside semantic reasoning rather than being assumed to emerge automatically from general-purpose multimodal representations.

6. Empirical evidence, ablations, and relation to adjacent rectification literatures

The empirical results reported for the FRM are substantial. For cross-dataset localization, the full system achieves average localization F1-score and IoU of 0.423 and 0.351, outperforming the second-best method PIM at 0.342/0.284. The paper also states that it improves over the recent MLLM-based baseline SIDA by 17.9 percentage points (F1). For image-level detection, it achieves average F1-score of 0.809, beating the next-best method by 6.7 percentage points (Zhang et al., 25 Aug 2025).

The ablation evidence is especially direct about the FRM’s importance. The paper reports that removing FRM causes major drops in both detection and localization: detection F1 falls from 0.809 to 0.676, localization F1 from 0.423 to 0.301, and IoU from 0.351 to 0.225. It further reports that removing analysis-informed feature gating also lowers performance, and that removing the MLLM proposal generator causes declines that are smaller than removing FRM. The paper therefore describes FRM as the cornerstone of the framework (Zhang et al., 25 Aug 2025).

The robustness evidence follows the same pattern. The FRM-based system is reported to be more robust to various perturbations, including brightness, contrast, noise, and compression, maintaining high performance where competing methods degrade rapidly. Qualitative results are described as showing cleaner, more accurate boundaries and fewer false positives, especially on challenging images without obvious semantic boundaries (Zhang et al., 25 Aug 2025).

The term “rectification” has a wider technical history, and the FRM occupies one specific branch of it. In document analysis, rectification usually denotes geometric dewarping. SalmRec uses self-adaptive multi-task fusion, inter-task feature aggregation, gating modules, and a Transformer decoder to predict a deformation field for distorted document images (Li et al., 9 May 2025). Marior separates Margin Removal Module and Iterative Content Rectification Module, with dense displacement flow prediction and adaptive iteration (Zhang et al., 2022). ForCenNet centers foreground elements, mask-guided transformer attention, and curvature consistency loss for document image rectification (Cai et al., 26 Jul 2025). In fingerprint forensics, rectification denotes dense distortion field estimation from a single distorted fingerprint image, improving rectified fingerprint matching (Guan et al., 2024, Guan et al., 2024). In long-video forensics, MSLoc performs coarse-to-fine temporal forensic rectification by combining a boundary-sensitive proposal generation module with an MLLM-based refinement module for temporal localization and explanation (Feng et al., 1 Jun 2026).

These neighboring uses show that “rectification” is a family resemblance term rather than a single technique. In the specific case of the Forensics Rectification Module, the object being rectified is not geometry, fingerprint skin deformation, or temporal boundaries alone. It is the semantic proposal itself: an initial MLLM hypothesis is reweighted, validated, and refined until its embeddings are consistent with multi-scale forensic evidence (Zhang et al., 25 Aug 2025).