RGB Histogram

The idea is to generate a 3-D RGB histogram (bins for R and G and B combinations) for a series of images so we can order them by similarity or search for images most similar to an existing image.

The code basically comes from:

http://www.pyimagesearch.com/2014/01/27/hobbits-and-histograms-a-how-to-guide-to-building-your-first-image-search-engine-in-python/

Files:

• chi_square_distance.sql: A SQL function to calculate the chi square distance between to histogram arrays.
• chi_square_distance.c: C version of the chi square distance calculation.
• Makefile: Makefile for the C shared library. After building:

$ sudo cp chi_square_distance.so $(pg_config --pkglibdir)

CREATE FUNCTION chi_square_distance(a double precision[], b double precision[])
RETURNS double precision
AS 'chi_square_distance', 'chi_square_distance'
LANGUAGE C STRICT;

• generate_photolog_histograms.py: A Python script to populate the photolog_histograms table.
• rgb_histogram.py: A Python script that builds a CSV of 3D RGB histogram distances for all photos in a directory.

Updating distances:

INSERT INTO photolog_histogram_distances
(a_id, b_id, bins, distance)
(SELECT a.id, b.id, 32,
   chi_square_distance(a.rgb_histogram, b.rgb_histogram)
 FROM photolog_histograms AS a
   INNER JOIN photolog_histograms AS b
      ON a.id < b.id AND a.bins = b.bins
 WHERE a.bins = 32 AND a.id || '-' || b.id NOT IN
 (SELECT a_id || '-' || b_id FROM photolog_histogram_distances
  WHERE bins = 32));

REFRESH MATERIALIZED VIEW all_histogram_distances;

Notes:

• generate_photolog_histograms.py took 45 minutes with 32 bins.
• rgb_histogram.py took a little over 29 minutes to do every pairwise distance twice (1-2 and 2-1) with 8 bins.
• chi_square_distance.sql took just under two hours to do every pairwise distance once with 8 bins.
• chi_square_distance.c took 79 seconds using 8 bin histograms and 450 seconds using 16 bit histograms. 32 bin histogram distances took 54 minutes to compute.