puzzle_set man page
puzzle_set_max_width, puzzle_set_max_height, puzzle_set_lambdas, puzzle_set_p_ratio, puzzle_set_noise_cutoff, puzzle_set_contrast_barrier_for_cropping, puzzle_set_max_cropping_ratio, puzzle_set_autocrop — set tunables for libpuzzle functions.
puzzle_set_max_width(PuzzleContext *context, unsigned int width);
puzzle_set_max_height(PuzzleContext *context, unsigned int height);
puzzle_set_lambdas(PuzzleContext *context, unsigned int lambdas);
puzzle_set_p_ratio(PuzzleContext *context, double p_ratio);
puzzle_set_noise_cutoff(PuzzleContext *context, double noise_cutoff);
puzzle_set_contrast_barrier_for_cropping(PuzzleContext *context, double barrier);
puzzle_set_max_cropping_ratio(PuzzleContext *context, double ratio);
puzzle_set_autocrop(PuzzleContext *context, int enable);
While default values have been chosen to be ok for most people, the puzzle_set_*() functions are knobs to fit the algorithm to your set of data and to your applications.
By default, pictures are divided in 9 x 9 blocks.
9 is the lambdas value, and it can be changed with puzzle_set_lambdas()
For large databases, for complex images, for images with a lot of text or for sets of near‐similar images, it might be better to raise that value to 11 or even 13
However, raising that value obviously means that vectors will require more storage space.
The lambdas value should remain the same in order to get comparable vectors. So if you pick 11 (for instance), you should always use that value for all pictures you will compute a digest for. puzzle_set_p_ratio()
The average intensity of each block is based upon a small centered zone.
The "p ratio" determines the size of that zone. The default is 2.0, and that ratio mimics the behavior that is described in the reference algorithm.
For very specific cases (complex images) or if you get too many false positives, as an alternative to increasing lambdas, you can try to lower that value, for instance to 1.5.
The lowest acceptable value is 1.0.
In order to avoid CPU starvation, pictures won't be processed if their width or height is larger than 3000 pixels.
These limits are rather large, but if you ever need to change them, the puzzle_set_max_width() and puzzle_set_max_height() are available.
The noise cutoff defaults to 2. If you raise that value, more zones with little difference of intensity will be considered as similar.
Unless you have very specialized sets of pictures, you probably don't want to change this.
By default, featureless borders of the original image are ignored. The size of each border depends on the sum of absolute values of differences between adjacent pixels, relative to the total sum.
That feature can be disabled with puzzle_set_autocrop(0) Any other value will enable it.
puzzle_set_contrast_barrier_for_cropping() changes the tolerance. The default value is 5. Less shaves less, more shaves more.
puzzle_set_max_cropping_ratio() This is a safe‐guard against unwanted excessive auto‐cropping.
The default (0.25) means that no more than 25% of the total width (or height) will ever be shaved.
Functions return 0 on success, and -1 if something went wrong.