Detects the visible app window in related screenshots, and then crops, pads, and outputs all of them at a consistent size with uniform spacing around the window - useful for a set of screenshots taken at slightly different sizes (e.g.: from using a free-hand screenshot tool).
- Window detection via ML background removal (rembg + U2-Net), luminance projection, background colour distance, edge detection, and contour analysis - results are intersected for best guess
- Consensus sizing: all output images share the same dimensions
- If filling is needed: uniform wallpaper-coloured border sampled from the original background
- Python 3.10+
- uv
uv run normaliser.pyuv run reads the inline script metadata and installs all dependencies (opencv-python, Pillow, numpy, rembg) into a temporary environment.
uv sync
uv run python normaliser.py- Click Add images to load one or more screenshots.
- Click Analyse images - detects the application window in each image. Click through the image list to verify.
- Adjust Padding (pixels added around the detected window) and Detect sensitivity if needed.
- Click Process and save - output files are saved as
<name>_normalised<ext>.
All five detection methods run on every image, and their results are intersected to produce a conservative, robust bounding box:
- ML(rembg/U2-Net) - removes the wallpaper background and derives the window bounds from the alpha mask.
- Luminance projection - Otsu threshold + row/column fraction arrays; finds the longest contiguous run above a threshold in each axis.
- Background colour distance - compares each pixel to the corner-sampled wallpaper colour.
- Edge/Hough-line detection - looks for long horizontal and vertical lines bounding the window.
- Contour analysis - finds the largest external contour after edge dilation.
Each method's bounding box is intersected (maximum of left/top edges, minimum of right/bottom edges), so that methods which expand too far toward the image boundary are clipped by the more conservative ones.
Run endpoints are also trimmed using a relative threshold: trailing rows/columns whose foreground density falls below 50% of the peak density inside the detected region are removed. This handles wallpaper figures or gradients that cross the absolute threshold but are clearly outside the window.
| Package | Purpose |
|---|---|
opencv-python |
Thresholding, edge detection |
Pillow |
Image I/O, resize, compositing |
numpy |
Array operations |
rembg |
ML background removal (U2-Net) |
MIT © 2026 Michael Champanis