Apply exp() to the Retinex output before normalization, restoring a
linear tone distribution. This improves contrast overall: light gray
content is preserved, and dark content stays dark even when the image
contains deep blacks. The previous normalization compressed all tones
toward the same gray level. Trade-off: shadows and unwanted background
variations are also more visible in some images.
Other:
- stretch toward white using histogram mode
- fall back to simple normalization on documents with large dark areas
- add bilateral denoising
Goal: improve precision of automatic document cropping by switching:
- from Douglas-Peucker algorithm (OpenCV's approxPolyDP) + a heuristic for documents missing a corner
- to an algorithm that looks for edges
* New quad algorithm: identify edges from contour orientation
* Performance optimization: reduce number of calls to trigonometric functions
* Performance: use a single threshold for live analysis
* Fix orientation of debug mask and compute it only if required
* Exclude quads that go out of the frame
Better differentiate color and grayscale documents:
- Look for colored pixels only in the intersection of the mask and quadrilateral
- Apply a white balance (grey world) to the document
- Exclude pixels with extreme luminance
- Erode segmentation mask
