An Evaluation of Popular Copy-Move Forgery Detection Approaches

I. INTRODUCTION

• CMFD detects content copied and pasted within one image, where shared source and target image conditions are exploited by many algorithms.Typical motives include hiding an element or emphasizing objects.Source: §I · p. 1
• The paper asks which CMFD methods work across differing image attributes, addressing prior comparisons that were typically limited to the algorithm under examination.It evaluates 15 feature sets in a joint pipeline and tests common CMFD scenarios because the forger’s processing is not known beforehand.Source: §I · p. 1

II. TYPICAL WORKFLOW FOR COPY-MOVE FORGERY DETECTION

• Block-based methods extract features from rectangular image regions, whereas keypoint-based methods extract features only from high-entropy regions.Both match similar features and filter spurious matches; clustered matches can be grouped by a shared transformation pattern.Source: §II · p. 2, Figure 2 · p. 2
• For block-based methods, the workflow filters matches using a threshold on the SATS-connected area; keypoint methods use homography-based clustering.The evaluation combined Amerini et al.’s clusters with RANSAC-estimated affine transforms and Pan and Lyu’s correlation map for pixel-wise keypoint results.Source: §II · pp. 2–3
• After feature extraction, the joint workflow matches feature vectors, removes pairs with too-small shift vectors, and clusters remaining matches that follow a common pattern.The final decision marks an image tampered when a connected region exceeds a threshold.Source: §II · p. 3

A. Block-based Algorithms

• Thirteen block-based features span moment-, dimensionality-reduction-, intensity-, and frequency-based categories.The evaluated set includes ZERNIKE, PCA, KPCA, DCT, and DWT; a DWT-plus-SVD approach was excluded after unreliable results in this setup.Source: A. Block-based Algorithms · p. 3
• The paper fixes block size at 16 × 16 for feature sizes that depend on block size; PCA and SVD feature sizes depend on image or block content, respectively.Source: B. Keypoint-based Algorithms · p. 3

B. Keypoint-based Algorithms

• SIFT and SURF are the two evaluated keypoint feature sets, with features extracted at high-entropy image regions rather than across tiled blocks.Fewer feature vectors reduce matching and postprocessing cost, but copied regions may be sparsely covered or missed when they have little structure.Source: B. Keypoint-based Algorithms · pp. 3–4

III. BENCHMARK DATA

• Existing datasets were considered limited for in-depth CMFD evaluation, including low-resolution examples and databases whose copy-move processing was limited to rotation and scaling.The MICC databases lacked source files, preventing added noise or other artifacts on copied regions.Source: §III · p. 4
• The benchmark uses 48 source images and 87 manually prepared, semantically meaningful snippets varying in size and content texture.A software framework applies common noise sources, JPEG artifacts, scaling, and rotation while generating pixel-wise ground truth.Source: §III · p. 4
• The forgeries target high-resolution consumer-camera images averaging about 3000×2300 pixels, with around 10% of pixels in tampered regions.The authors define realism by copied-pixel count, boundary treatment, and snippet content.Source: §III · p. 4
• The test cases vary in copied-region texture, contrast, self-similarity, and detail, including low-contrast cat babies and a high-detail jellyfish scene.Source: §III · p. 5, Figure 5 · p. 5

IV. ERROR MEASURES

• The evaluation measures both whether an image is tampered and how accurately its copied regions are identified at pixel level.Partially transparent or otherwise mixed boundary pixels are not well-defined for CMFD evaluation and are excluded.Source: §III · p. 4, §IV · p. 5

A. Metrics

• Precision is the probability that a detected forgery is real, while recall is the probability that a forged image is detected.At pixel level, the same measures use correctly detected, falsely detected, and missed forged pixels.Source: A. Metrics · pp. 5–6
• F1 combines precision and recall into a single value.The paper also reports F1 in its tables.Source: A. Metrics · p. 6

B. Protocol

• Pixelwise evaluation gives a more detailed view of detected copied regions than image-level classification, but requires a careful definition of which pixels count as copied.The ground truth draws boundaries even when source and target intensities are indistinguishable, excludes smeared or partially transparent pixels, and stays clean under postprocessing.Source: B. Protocol · p. 6

V. EXPERIMENTS

• The roughly 4,700 variants cover different snippet scales, rotation angles, compression rates, and combinations, with experiments assessing image-level and pixelwise detection.The pixelwise series provides a more detailed assessment of feature discrimination; source code and full result tables were also made available.Source: §V · p. 6

A. Threshold Determination

• Thresholds were mostly fixed across methods for fairer comparison, while feature-specific thresholds were manually tuned to the benchmark.For block features, τ2 was chosen by optimizing image-level F1 across JPEG quality levels 100 to 70; keypoint methods used τ2 = 4 correspondences.Source: A. Threshold Determination · pp. 6–7

1) Plain Copy-Move:

• 13 of 15 features achieved 100% recall on plain copies, but only four methods exceeded 90% precision.Even ideal conditions produced false alarms, attributed mainly to diverse image challenges and the higher chance of false-positive matches in large images.Source: 1) Plain Copy-Move · p. 7, Table II · p. 7

2) Robustness to Gaussian noise:

• Under Gaussian noise, DCT retained notably high recall; PCA, SIFT, SURF, and HU also maintained good recall at high noise levels.Seven of 15 features fell below 50% recall, while precision decreased only slightly; SURF offered a good precision-recall balance, followed by PCA.Source: 2) Robustness to Gaussian noise · p. 7

3) Robustness to JPEG compression artifacts:

• At JPEG quality levels considered reasonable for real-world forgeries, SURF and SIFT precision remained stable, while block-based methods’ precision gradually fell to 0.5.KPCA, DCT, ZERNIKE, BLUR, and PCA each maintained recall of at least 90%.Source: 3) Robustness to JPEG compression… · p. 8, Figure 8 · p. 8

4) Scale-invariance:

• Most features degraded quickly under scaling, but KPCA, ZERNIKE, LUO, DWT, DCT, and PCA handled moderate scaling; SIFT and SURF were most stable across the full tested range.The copied region was scaled from 91% to 109% in 2% steps, with additional tests at 50%, 80%, 120%, and 200%.Source: 4) Scale-invariance · p. 8

5) Rotation-invariance:

• For rotation, ZERNIKE had the best precision; BRAVO and ZERNIKE had consistently good recall, while SURF performed best at higher rotation angles.KPCA and LUO also performed strongly for small rotations; FMT, LIN, HU, and BLUR were not well suited to rotation variation.Source: 5) Rotation-invariance · p. 8

6) Robustness to Combined Transformation:

• For combined transformations, DCT and ZERNIKE gave the best recall under small variations, while SIFT and SURF had the best precision and very good recall under large rotations and scaling.At the fourth step, SURF and SIFT were almost on a par with ZERNIKE; several features had recall below 50%.Source: 6) Robustness to Combined Transformation · p. 8, Table III · p. 9

C. Detection at Pixel Level

• Pixel-level evaluation checks matched-block centers against pixelwise ground truth, excludes boundary pixels, and reports postprocessed keypoint matches.Because keypoint matches are sparse, their pixelwise performance cannot be directly compared fairly with block-based methods.Source: C. Detection at Pixel Level · p. 9

1) Plain Copy-Move:

• The pixel-level plain-copy baseline used unmodified copy-pastes, with thresholds calibrated for comparable detection performance across methods.The cited passage identifies the baseline results in Table III but provides no numerical comparisons.Source: 1) Plain Copy-Move · p. 9, Table III · p. 9

2) Robustness to Gaussian noise:

• DCT, SIFT, and SURF provided the best pixel-level recall under Gaussian noise, and DCT also outperformed all other methods in precision.CIRCLE, BLUR, BRAVO, SVD, and HU were very noise-sensitive; PCA, ZERNIKE, KPCA, and DWT deteriorated more gradually.Source: 2) Robustness to Gaussian noise · p. 9, Figure 10 · p. 9

3) Robustness to JPEG compression artifacts:

• SIFT, SURF, KPCA, DCT, PCA, and ZERNIKE are relatively resilient to JPEG compression; DWT is slightly less resilient.DCT's robustness was expected because its features use the transform that also underlies JPEG.Source: 3) Robustness to JPEG compression… · p. 9

6) Robustness to Combined Transformation:

• SURF and SIFT perform considerably well under combined scaling, rotation, and JPEG compression, followed by ZERNIKE, DCT, KPCA, and DWT.The authors treat these experiments as an indicator that single-transformation findings approximately transfer to joint transformations, while noting the full parameter space is infeasible to test.Source: Figure 11 · p. 10, 6) Robustness to Combined Transformation · p. 10

4) Scale-invariance:

• For scaling differences greater than 9%, SIFT and SURF are the only feature sets with acceptable precision and recall.For small scaling changes, seven feature sets show invariance, with DWT, KPCA, ZERNIKE, PCA, and DCT the best-performing block-based methods.Source: 4) Scale-invariance · p. 10

5) Rotation-invariance:

• SIFT and SURF have the most stable recall under rotation; among block-based methods, ZERNIKE, BRAVO, and LUO perform best for larger rotations.FMT and CIRCLE also perform very well at 180° rotation.Source: 5) Rotation-invariance · p. 10

D. Detection of Multiple Copies

• With g2NN matching, DCT, DWT, KPCA, PCA, and ZERNIKE can benefit significantly, while SIFT and SURF drop considerably on randomly selected small copy regions.The authors attribute the keypoint methods' drop mainly to too few matched keypoints in those regions; high-entropy copy-region selection is suggested as a possible improvement, not evaluated.Source: Table IV · p. 10, D. Detection of Multiple Copies · p. 10

E. Downsampling: Computational Complexity versus Detection Performance

• Downsampling considerably reduces performance for all features, with recall more affected than precision; KPCA, ZERNIKE, DWT, PCA, DCT, LUO, FMT, and BRAVO perform relatively well.Detection parameters were globally fixed, so the results are presented as a conservative bound on theoretically best performance.Source: E. Downsampling: Computational… · p. 11

F. Resource Requirements

• Block-based feature storage can reach several gigabytes on large images; DCT and DWT peak above 20GB in the implementation, while 8GB sufficed for most feature sets.The theoretical minimum is more than doubled in the implementation; DCT and DWT feature sizes depend on block size, which can be reduced in practice.Source: F. Resource Requirements · p. 11
• Among block-based methods, intensity features compute quickly, while BLUR, DCT, and KPCA are most costly; PCA, FMT, and ZERNIKE are also relatively demanding.The reported computation times are averages per image, split into feature extraction, matching, and postprocessing.Source: F. Resource Requirements · p. 11
• Keypoint-based methods excel in computation time and memory consumption because they typically extract an order of magnitude fewer features than the number of image blocks.Their feature size is relatively large, but subsequent processing is lightweight.Source: F. Resource Requirements · p. 11

G. Qualitative Results

• Under JPEG quality 70, SURF misses two of three copied passengers while ZERNIKE detects all three but also marks many sky-region false positives.For a 20° rotation, both methods accurately detect copied regions; in a combined-transformation example, background symmetries produce many SURF false positives while ZERNIKE detects the copies.Source: G. Qualitative Results · p. 12

H. Results by Image Categories

• Object-class performance differences are relatively small, but texture matters: SIFT and SURF excel on rough regions, while block-based methods often outperform them on smooth regions.Structure-category performance lies between those two texture extremes.Source: H. Results by Image Categories · p. 12

VI. DISCUSSION

• DCT, PCA, KPCA, ZERNIKE, and DWT outperform other block features in robustness tests, while SIFT and SURF excel for copies with large rotation or scaling.Keypoint methods can miss low-structure regions and generate false positives on highly self-similar content.Source: §VI · p. 13
• Downsampling reduces overall performance, although the strongest block-based feature sets remain relatively stable at a lower performance level.The authors note that full-sized block-based processing can exceed available resources and quantify the loss when images are not processed at original size.Source: §VI · p. 13, §VII · p. 14
• In the smaller-scale multiple-copy evaluation, DCT, DWT, KPCA, PCA, and ZERNIKE profited most from improved postprocessing.The matching and filtering steps were adapted to use g2NN and consider the n best-matching features.Source: §VI · p. 13
• A practical system can pair efficient keypoint-based screening with block-based close examination, for which the authors recommend ZERNIKE features.The authors report that block-based methods are more accurate when added noise or transformations are small.Source: §VI · p. 13
• Final recommendations depend on the detection scenario; resampling or double-JPEG analysis may complement CMFD when their traces are available.The authors describe resampling traces for rescaled or rotated copies under moderate JPEG compression and possible double-JPEG artifacts for regions copied within JPEG images.Source: §VI · p. 13

VII. CONCLUSION

• The evaluation framework combines 48 realistically sized base images, 87 copied snippets, and software for replaying controlled, realistic-looking forgeries.The study evaluates both per-image tampered/original classification and per-pixel performance.Source: §VII · p. 13
• SIFT-based keypoint detection combines low computational load with good performance, while ZERNIKE is recommended among the standout block features for its relatively small memory footprint.The other standout block-based features are DCT, DWT, KPCA, and PCA.Source: §VII · pp. 13–14

VIII. CONTENTS OF THE SUPPLEMENTAL MATERIAL

• The supplement also provides F1 scores for experiments reported in the main paper, alongside threshold-selection and database details.F1 combines precision and recall in a single score.Source: §VIII · p. 15

IX. THRESHOLD DETERMINATION FOR THE MINIMUM NUMBER OF CORRESPONDENCES

• τ2 filters connected matched-correspondence clusters below a minimum size, so it is set separately for each feature type to accommodate different match densities.A high threshold could discard small disconnected detections, while a low threshold could increase sensitivity to noise in dense matches.Source: §IX · p. 15
• The authors searched τ2 values from 50 to 1000 on clean and moderately JPEG-compressed images, selecting the best value for each method.JPEG quality ranged from 100% to 70% in steps of 10% to reduce overfitting to clean data.Source: §IX · p. 15
• JPEG artifacts decrease the average optimum τ2 to almost 50% of its plain-copy value.The thresholds were determined using per-image F1 scores on plain copies versus mixed clean and JPEG-compressed images.Source: §IX · p. 15, Table VI · p. 15
• Method-specific τ2 values outperform fixed τ2 = 1000 by a large margin on combined transformations, even without dedicated fine-tuning.The combined case uses 2° rotation, 101% upscaling, and JPEG quality 80.Source: §IX · p. 16, Table VIII · p. 16

X. RESULTS WITH F1 SCORE

• Recall mainly shapes F1 scores across noise, JPEG compression, scaling, rotation, and downsampled-image scenarios.At image level, precision and recall differences for JPEG compression are evened out; at pixel level, recall differences remain prominent.Source: §X · p. 17

XI. POSTPROCESSING OF KEYPOINT-BASED METHODS

• Keypoint postprocessing clusters matched pairs, estimates affine transformations with RANSAC, then combines thresholded correlation maps into copied-region masks.The maps use 5 × 5-pixel normalized correlation windows, a 7 × 7 Gaussian smoothing kernel, and a 0.4 binarization threshold.Source: §XI · pp. 17–18

XII. DATABASE CATEGORIES

• The supplemental category analyses subdivide the dataset by image content to help explain feature-descriptor performance differences.The authors use both object and texture categories.Source: §XII · p. 18

A. Categorization by Object Classes

• For plain copies, object categories perform comparably at pixel level, while most feature sets perform best on living and nature at image level.The living category contains 5 test cases and man-made contains 24; category sizes vary, but evaluation is computed within each category.Source: A. Categorization by Object Classes · p. 18, Table X · p. 18
• Across postprocessing scenarios, the best pixel-level results sometimes occur in man-made and mixed categories, while several feature sets do better on nature than living.The authors suggest clearer structure in man-made objects as a possible explanation.Source: A. Categorization by Object Classes · p. 18, Figure 18 · p. 19, Figure 19 · p. 19

B. Categorization by Texture

• Texture categories were assigned from artists’ fuzzy descriptions rather than quantitative measurements, reflecting the authors’ real-world-forgery perspective.Smooth and rough are approximate low- and high-entropy distinctions; structured content usually has pronounced edges and corners.Source: B. Categorization by Texture · pp. 18–19
• SURF and SIFT perform best on rough textures, while block-based methods often have an advantage on smooth textures.Texture groups are smooth, rough, and structured; copied-area size was not distinguished in this evaluation.Source: B. Categorization by Texture · p. 19, Figure 19 · p. 19, Figure 18 · p. 19

• The database contains 48 base images and 87 prepared snippets, which can be spliced with postprocessing artifacts and paired with pixelwise ground truth.The framework can iterate over parameters such as JPEG quality to create image series, and its separate components support adding other manipulation scenarios.Source: §XIII · p. 20
• Reference examples show plain-copy manipulations and ground truth for smooth, rough, and structure categories, with white marking copy-source or copy-target regions.The examples vary in copied-region size and detail, and the authors note that category assignment can be debatable.Source: §XIII · p. 21, Figure 25 · p. 23, Figure 26 · p. 24
• The framework allows snippets to be inserted at arbitrary positions, so it can also produce semantically meaningless forgeries.The authors note that rotation and resampling can make image content naturally implausible.Source: §XIII · p. 21