ADCD-Net: Document Forgery Localization

• ADCD-Net is a document forgery localization model that combines adaptive DCT feature modulation with hierarchical content disentanglement to identify tampered areas in JPEG documents.
• The model fuses multi-scale RGB and JPEG-domain features using an alignment score to effectively manage distortions such as compression, resizing, blur, and noise.
• Empirical evaluations show ADCD-Net outperforms competitors by improving F1 scores by ~20% and reducing text–background bias through pristine prototype estimation.

ADCD-Net is a document image forgery localization model for detecting tampered regions in JPEG document images by producing a dense binary mask of forged versus pristine pixels. It was introduced as a robust, document-specific architecture that combines RGB-domain and JPEG-domain evidence, explicitly addressing the sensitivity of DCT traces to block-grid misalignment, the text–background disparities characteristic of document imagery, and the need for robustness under recompression, resizing, blur, and noise. Architecturally, it is a U-shaped, dual-decoder network built around adaptive DCT feature modulation, hierarchical content disentanglement, and pristine prototype estimation (Wong et al., 22 Jul 2025).

1. Nomenclature and scope

In current arXiv usage, ADCD-Net most precisely denotes the document forgery localization model introduced in 2025, titled “ADCD-Net: Robust Document Image Forgery Localization via Adaptive DCT Feature and Hierarchical Content Disentanglement” (Wong et al., 22 Jul 2025). A recurrent source of confusion is the existence of several similarly named but unrelated models: “ADC-Net” for automated dispersion compensation in optical coherence tomography (Ahmed et al., 2022), “ACDnet” for real-time action detection on edge devices (Liu et al., 2021), “ADCNet” for radar perception from raw ADC data (Yang et al., 2023), SEMI-SuperYOLO-NAS as an Automated Defect Classification and Detection framework in semiconductor inspection (Chen et al., 2024), and CDAD-Net for across-domain generalized category discovery (Rongali et al., 2024).

Within the document forensics literature, the term therefore refers to a model specialized for document image forgery localization, not to OCT restoration, video action detection, radar distillation, semiconductor defect detection, or category discovery. That disambiguation matters because the 2025 ADCD-Net is defined by a distinct forensic setting: JPEG document input, DCT-domain priors, OCR-derived background modeling, and dense tamper-mask prediction.

2. Problem formulation and forensic setting

Document image forgery localization seeks to identify tampered content within document images at pixel level or region level by producing a binary mask of forged versus pristine pixels. The task differs materially from natural-image forgery detection because document images exhibit unusually uniform backgrounds, highly structured text with sharp contours, and consistent textures. Typical manipulations replace or inpaint small text spans that blend seamlessly with surrounding content, so methods that rely on scene semantics, camera fingerprints, or broad texture cues can miss small tampered regions or overreact to pristine text (Wong et al., 22 Jul 2025).

ADCD-Net is formulated for JPEG document inputs. The model takes an image $X \in \mathbb{R}^{H \times W \times 3}$ together with JPEG-domain signals: quantized Y-channel DCT coefficients $X_{dct} \in \mathbb{Z}^{H \times W}$ and the JPEG quantization table $X_{qt}$. Its output is a dense binary mask $\hat{Y} \in \{0,1\}^{H \times W}$ indicating tampered pixels. The design is driven by three stated challenges: the sensitivity of DCT features to $8 \times 8$ block-grid alignment, the need for robustness across distortions such as Gaussian noise, Gaussian blur, downscaling, and single or multiple JPEG compression, and the text–background bias caused by the different frequency and intensity statistics of text and blank background.

The JPEG-domain component is central because DCT artifacts are powerful forensic cues, yet brittle under misalignment. For an $N \times N$ block $B(x,y)$ with $N=8$, the standard 2D DCT is

$$F(u,v) = \alpha(u)\alpha(v)\sum_{x=0}^{N-1}\sum_{y=0}^{N-1} B(x,y)\cos\left[\frac{(2x+1)u\pi}{2N}\right]\cos\left[\frac{(2y+1)v\pi}{2N}\right],$$

where $\alpha(0)=\sqrt{1/N}$ and $X_{dct} \in \mathbb{Z}^{H \times W}$0 for $X_{dct} \in \mathbb{Z}^{H \times W}$1. ADCD-Net uses the quantized coefficients derived from this process. Cropping, resizing, or pixel shifts with offsets not divisible by $X_{dct} \in \mathbb{Z}^{H \times W}$2 disrupt the original block alignment, scrambling the relationship between DCT coefficients and image content. The model’s adaptive DCT mechanism is explicitly intended to modulate reliance on these cues when they become unreliable.

3. Architecture and internal representations

ADCD-Net is a U-shaped, dual-decoder network specialized for document images. Its encoder is an adaptive RGB–DCT encoder $X_{dct} \in \mathbb{Z}^{H \times W}$3, where $X_{dct} \in \mathbb{Z}^{H \times W}$4 extracts multi-scale RGB-domain forensic features and $X_{dct} \in \mathbb{Z}^{H \times W}$5 extracts multi-scale DCT-domain features while also predicting a global alignment score $X_{dct} \in \mathbb{Z}^{H \times W}$6. Fusion occurs at each scale through

$X_{dct} \in \mathbb{Z}^{H \times W}$7

so the DCT contribution is gated by a single image-level alignment estimate rather than injected unconditionally (Wong et al., 22 Jul 2025).

The DCT encoder uses the Frequency Prior Head (FPH) from DTD. The RGB encoder and both decoders are based on Restormer blocks. The alignment score is predicted from the last-level DCT feature $X_{dct} \in \mathbb{Z}^{H \times W}$8 by a classification head attached to the final block of $X_{dct} \in \mathbb{Z}^{H \times W}$9. Supervision is binary: aligned samples are labeled $X_{qt}$0, while non-aligned samples are created by augmentations that break the grid, including random resizing, cropping, and pixel shifting with $X_{qt}$1, and labeled $X_{qt}$2.

A defining component is the Hierarchical Content Decoupling (HCD) module $X_{qt}$3. At each scale $X_{qt}$4, fused features are split into content and forgery branches:

$X_{qt}$5

This is implemented by lightweight per-scale MLPs. The purpose is to reduce text–background bias by disentangling document content from tamper-specific traces. Two reconstruction pathways enforce this separation. One uses $X_{qt}$6, while the other uses $X_{qt}$7. The spatial shuffling breaks content alignment in the forgery branch, forcing document structure to reside primarily in $X_{qt}$8.

The two decoders serve different roles. The reconstruction decoder $X_{qt}$9 reconstructs both the original RGB image $\hat{Y} \in \{0,1\}^{H \times W}$0 and DCT coefficients $\hat{Y} \in \{0,1\}^{H \times W}$1, rather than RGB alone. The localization decoder $\hat{Y} \in \{0,1\}^{H \times W}$2 processes forgery features and produces a multi-scale representation for segmentation. Before the final segmentation head $\hat{Y} \in \{0,1\}^{H \times W}$3, the penultimate forgery feature is modulated by the Pristine Prototype Estimation (PPE) mechanism. Using an OCR-derived background mask $\hat{Y} \in \{0,1\}^{H \times W}$4 from CRAFT, ADCD-Net computes a pristine background prototype at each scale:

$\hat{Y} \in \{0,1\}^{H \times W}$5

Per-pixel cosine similarity to this prototype is then

$\hat{Y} \in \{0,1\}^{H \times W}$6

Two MLPs convert these similarity maps into scale and bias tensors, and final prediction is produced as

$\hat{Y} \in \{0,1\}^{H \times W}$7

The stated intuition is that pixels whose forgery features resemble the pristine prototype are down-weighted, while anomalous ones are emphasized.

4. Objectives, supervision, and training regime

The complete training objective is

$\hat{Y} \in \{0,1\}^{H \times W}$8

The alignment loss is binary cross-entropy on the predicted alignment score:

$\hat{Y} \in \{0,1\}^{H \times W}$9

The reconstruction loss is an $8 \times 8$0 loss applied to both reconstruction paths:

$8 \times 8$1

The forgery localization loss combines cross-entropy and Lovasz loss on the final mask. The within-image supervised contrastive loss $8 \times 8$2 operates over multi-scale forgery features, with sampling used to control computation and class imbalance. Its stated role is to amplify separation between pristine and tampered pixels and to weaken text–background bias by pulling pristine text closer to pristine background while pushing tampered text apart (Wong et al., 22 Jul 2025).

Training uses DocTamper with masks, and evaluation is cross-domain on the Test, FCD, and SCD sets defined in DocTamper. Robustness-oriented augmentations include random resizing or downscaling, cropping, pixel shifting with $8 \times 8$3, Gaussian blur, Gaussian noise, and JPEG recompression, both single and multiple. Initialization uses DocRes weights for Restormer blocks. The optimizer, learning-rate schedule, batch size, and epochs are not specified in the main paper. The model has 23M parameters, reported as the smallest among compared methods in DocTamper’s multi-JPEG setting. FLOPs and FPS are not reported.

5. Empirical behavior and ablation findings

ADCD-Net is reported to consistently outperform state-of-the-art methods by 20.79\% averaged over 5 types of distortions, with the second-best competitor identified as TruFor in the detailed synthesis. The robustness evaluation covers Gaussian noise, Gaussian blur, downscaling, single JPEG compression, and multiple JPEG compression. In the multi-JPEG setting with quality factor $8 \times 8$4, ADCD-Net achieves average $8 \times 8$5, slightly above PS-Net at approximately $8 \times 8$6, while using 23M parameters. On pristine documents from the ICDAR 2023 Detecting Tampered Text in Images Tianchi dataset, it yields false alarm rate $8 \times 8$7 and the highest $8 \times 8$8 among compared methods (Wong et al., 22 Jul 2025).

The paper places particular emphasis on stress tests for DCT-grid disruption. Under small cropping (2\%), 1-pixel shift, and resize to 98\%, DTD suffers large $8 \times 8$9 drops of approximately 68–73\%, whereas ADCD-Net shows moderate declines of approximately 22–26\%. This is attributed to adaptive gating: when DCT becomes unreliable, the model can fall back to RGB-side cues.

Ablation studies isolate the contribution of the main components. Relative to the baseline, adding DCT features yields +$N \times N$0 $N \times N$1, adding within-image contrastive loss yields +$N \times N$2, and adaptive DCT modulation yields +$N \times N$3. Reconstructing both RGB and DCT, rather than RGB alone, improves performance by +$N \times N$4. PPE improves average performance when combined with HCD and adaptive DCT fusion, whereas using PPE without HCD degrades performance because prototype estimation from background-only aggregation becomes biased.

The bias analysis uses the cosine-similarity bias $N \times N$5 between background–pristine-text and tampered-text–pristine-text clusters. TruFor exhibits the highest bias; DTD reduces bias via DCT cues; ADCD-Net without HCD improves $N \times N$6 further through contrastive supervision; and ADCD-Net with HCD achieves the best $N \times N$7, which the authors interpret as evidence of effective mitigation of text–background bias. Qualitative reconstructions are described as showing that content features recover document layout in RGB and DCT channels, while forgery features retain characteristic noise and artifact patterns.

6. Robustness profile, limitations, and practical deployment

The model’s robustness claim is not that DCT evidence becomes invariant, but that its contribution is adaptively calibrated. The alignment score tends to be higher under JPEG compression and downscaling, and lower under crop, shift, resize, blur, and noise. A plausible implication is that ADCD-Net functions as a conditional forensic router: it leverages compression traces when the block structure remains informative and suppresses them when grid misalignment or signal corruption would otherwise inject error (Wong et al., 22 Jul 2025).

The paper also states several limitations. Extreme misalignment or heavy geometric transforms that erase compression evidence may still reduce confidence and sensitivity, forcing the model to rely on RGB cues only. Severe compression or very low-resolution scans can homogenize subtle textual edits. Complex layouts such as tables, stamps, seals, and handwriting may challenge OCR-based background segmentation, weakening pristine prototype estimation. Non-text tampering in dense graphics regions may benefit less from background-based prototypes, and performance then depends on how well HCD separates content from forgery traces.

The inference pipeline is straightforward but JPEG-aware. A JPEG image $N \times N$8 is loaded; the JPEG header is parsed to obtain $N \times N$9 and $B(x,y)$0; the RGB and DCT encoders are run jointly; $B(x,y)$1 is predicted; features are fused across scales; HCD splits them into content and forgery streams; the localization decoder produces multi-scale forgery features; OCR via CRAFT yields $B(x,y)$2; pristine prototypes and similarity maps are computed; and the final segmentation head predicts the tamper mask. The model outputs a probability map, and threshold $B(x,y)$3 is described as a common default. For stricter false-alarm control on pristine documents, thresholds in the $B(x,y)$4–$B(x,y)$5 range or validation-based calibration are suggested. Small connected components can be filtered by area to reduce spurious detections.

Implementation resources are public: the reported repository is https://github.com/KAHIMWONG/ACDC-Net, with code, data preparation scripts, and checkpoints referenced in the paper. The naming discrepancy between ADCD-Net in the paper title and ACDC-Net in the repository is therefore another practical point of clarification, but the underlying model described in the literature is the document-forensics system defined by adaptive DCT gating, hierarchical content disentanglement, and pristine prototype modeling.