M2Image forensicsMicroscopyLayer 1 (Deterministic)

Panel Overlap

Finds a field of view reused across the panels of a microscopy figure to stand in for different experimental conditions. Panels are compared by geometrically verified feature matching and by masked cross-correlation, with printed labels masked out so panels are flagged for shared image content rather than a shared caption.

Technical description

Splits the image into its four quadrants and into top, bottom, left, and right halves, then compares the six quadrant pairs and two half pairs with two detectors. ORB keypoints and binary descriptors are matched under Lowe's ratio test (0.75), and a RANSAC homography is fit; a pair passes on geometry when the inlier ratio is at least 0.5 with at least ten inliers and the homography is effectively affine, tested by requiring the perspective influence |h31| * width + |h32| * height below 0.1 so that a spurious projective fit on coincidental matches is rejected. In parallel, the masked normalized cross-correlation NCC = sum(f g) / sqrt(sum(f^2) sum(g^2)) is computed over pixels both regions agree are content, not text. A pair is accepted when the larger of the inlier ratio and the NCC exceeds its threshold (0.5 and 0.85).

How it works

Layer 1 (deterministic): masks printed text out of each region with an Otsu connected-component analysis, detects ORB keypoints in the content, matches them under Lowe's ratio test, and fits a RANSAC homography. A pair is flagged on geometry when a large fraction of matches are inliers and the recovered transform is affine, or on appearance when the masked cross-correlation is high. No accepted pair scores 0; one pair scores 2.0 + 2 * max_ncc and several score 3.0 + 0.5 * n_pairs, both capped at 5.0. Each pair yields a source and a duplicate finding with bounding boxes and a colour-coded difference overlay.

Why this matters

Reusing one image to represent several experiments is the most prevalent image-integrity problem in the published record; a survey of more than twenty thousand papers found problematic figures in 3.8 percent of them, with duplication the dominant category. A genuine multi-panel figure shows a distinct field of view per condition, so when the same capture, possibly shifted, rotated, flipped, or rescaled, appears in two panels, it is strong evidence of fabrication. Confirming the reuse from the pixels, with both a geometric and a photometric test, makes the detection robust to the editing a manipulator applies.

Score thresholds

0-1: No region reuse detected across quadrants or halves
2-3: One overlapping pair or several weak matches; a similar specimen can produce moderate similarity
4-5: One near-exact duplicate or multiple overlapping pairs, consistent with a field of view reused across conditions

Limitations

The screen reasons at the scale of quadrants and halves, so a small duplicated insert that does not align with those partitions can be missed. ORB struggles on low-texture content such as a uniform fluorescence field, leaving only the correlation detector. Genuinely similar specimens, periodic structures, and tiling patterns can raise the correlation without manipulation. The text mask assumes dark characters on a lighter background. The affine restriction will not flag a region strongly perspective-warped before pasting. General block-based copy-move anywhere in the image is handled by the clone-detection indicator I6; M2 specialises in panel-to-panel field-of-view reuse with geometric and correlation confirmation, and the two corroborate each other.