Use multiple screening signals, not one score
The browser screening workflow can combine exact-duplicate checks, perceptual similarity, copy-move and patch-duplicate evidence, cross-panel reuse checks, retouching or splicing signals, image metrics and JPEG-compression clues. A classical comparison mode also exposes similarity measures such as SSIM, perceptual hash and MSE. Different signals answer different questions, so one high or low score should not be treated as proof by itself.
Review matched regions in context
Candidate findings are presented in a reviewer-oriented evidence workspace with source context, panel labels and matched-region views. Geometry-based checks can use Local Engine OpenCV support when available. Optional experimental model-based evidence may also be available in some configurations, but it should be treated as corroborating evidence rather than a replacement for direct visual review.
A short review workflow
- Import the image set or manuscript PDF and confirm the source panels you want to screen.
- Run the available browser or classical checks and review the strongest candidate findings first.
- Open matched regions against their original images and check whether the similarity has a legitimate explanation.
- Record the evidence and export the available detector or review data when you need a reproducible follow-up.
Output you can keep
Keep a reviewer trail of the images, candidate findings and evidence you examined. The current workspace can export detector or analytics data as CSV for downstream review; use the source images alongside those records so a finding can be independently checked.
Interpretation limits
- The tool flags image-integrity signals; it does not determine research misconduct, author intent or publication eligibility.
- Repeated structures can be scientifically legitimate, and compression, resizing or figure assembly can affect similarity metrics.
- Weak or absent detector evidence does not prove that an image is unaltered.
- Any high-impact finding should be confirmed against original data, acquisition records and the scientific context.
