Automatic straightening

[akaihola]

Resources:

• imageproc::hough::detect_lines(), an implementation of the Hough transform
• rust-hough-transform by @maxbeutel
• Hough Transform Example (from @maxbeutel as well), understandable my math dummies

[akaihola]

The Hough Transform finds lines quite beautifully and quickly.

You can try by first turning an image into full black and white (but still RGB) using ImageMagick:

convert -threshold 90% original.png -type TrueColor -define png:color-type=2 black-and-white.png

Then run my branch of rust-hough-transform on the image:

cargo run --release black-and-white.png space.png line.png 1 1 1200

(replace the value 1200 of houghspace_filter_threshold with something like 80–90% of the max accumulator value as shown in terminal output from the command)

You can now inspect line.png for the strongest lines detected, and space.png for the raw accumulator output from the Hough Transform.

With 1° resolution, this takes about a second on my laptop for a 1200x1600 image.

[akaihola]

So based on lines detected by the Hough Transform, we could easily correct the rotation of a document based on the angles of strongest lines.

But we could also do a very basic "perspective correction", e.g. to straighten a bent book page:

• divide the image to four quadrants, i.e. 2 by 2 quarter images
• do the Hough transform to each quadrant separately
• get the angles of strongest horizontal and vertical lines
• assume those angles at the center point of each quadrant
• draw four curves, each going through two horizontally or vertically adjacent center points of quadrants and following the angle at those points – this could be done with e.g. Bezier curves
• let those curves form a "wavy" rectangle and use it as the guide for correcting the image
• at the simplest, a meshed curvy rectangle could be used to do a piecewise correction
• but maybe a more sophisticated smoother correction could be done