How It Works
Two detection engines working in parallel to catch image manipulation.
1

Upload an Image

Drag and drop or browse for any PNG, JPG, or WebP image. The file is sent to the backend where its signature bytes are verified to make sure it's actually an image before anything runs.

No file extension tricks get through. ForgeFind reads the raw binary header, not the filename.
2

Dual-Engine Analysis

Your image is processed by two independent detection systems at the same time. Each one targets a different kind of forgery.

Splicing Detection

PyTorch · U-Net · ResNet34 encoder

A deep learning segmentation model trained to identify regions of an image that were pasted in from a different source. It outputs a pixel-level mask highlighting suspicious areas.

  • Resizes input to 256x256 for inference
  • Upscales the probability map back to original resolution
  • Applies morphological cleanup to remove noise
  • Filters small blobs by confidence threshold
  • Rejects masks covering more than 25% of the image

Copy-Move Detection

OpenCV · SIFT · FLANN · RANSAC

A classical computer vision pipeline that finds regions duplicated within the same image. It extracts keypoints, matches them, and uses geometric verification to confirm cloned areas.

  • SIFT extracts scale-invariant feature descriptors
  • FLANN matches features with a ratio test (0.70)
  • Filters out spatially close matches (same region)
  • RANSAC fits a homography to verify geometric consistency
  • Rejects overlapping bounding boxes as false positives
3

Results Visualization

The original image is drawn to a canvas, and both detection outputs are layered on top. You can toggle between views to examine each finding independently or see everything at once.

Original Image

The unmodified upload with no overlays.

Noise Mask

Red overlay showing pixel regions the U-Net flagged as spliced. White mask pixels become semi-transparent red.

Clone Detection

Green bounding boxes around matched copy-move regions identified by the SIFT pipeline.

Overall

Both the red mask and green boxes together on a single view.

4

Confidence Scoring

Each engine produces its own score. The splicing confidence is the average probability across all flagged pixels. Copy-move detection is binary: if SIFT finds a verified clone, it reports 98% confidence.

The overall score takes the higher of the two, since either type of manipulation is enough to flag the image.

Score Ranges

0 - 30%  Low Risk. No meaningful manipulation detected.
31 - 70%  Medium Risk. Some suspicious regions found, but inconclusive.
71 - 100%  High Risk. Strong evidence of tampering.
5

Cleanup

Uploaded images and generated masks are stored temporarily on the server. A background task runs every 10 minutes and deletes any files older than 10 minutes. Nothing is kept permanently.