Image color analysis for your ecom site
22/04/2010Recently we’ve been working rather hard on a new look and complete re-launch of the underlying code for the whole of crookedtongues.com We’re a python shop these days so its running on django with a host of apps and the best part of over 100,000 lines of code.
One of the things that I was tasked with was the data migration and import and I’ll write about that post launch, what I fancy getting out now is how I’ve worked on adding color codes automagically to the store’s entire product database. More or less, its a little rough round the edges but gives fairly good results on a lot of our products. First off you can find the source code in a raw (read that as ‘I’m still working on it personally so take it as it comes’) way over at github. As a taste of what it does, here’s a couple of existing product shots:
It works quite well for colors that fall into the center areas of the color wheel. For my analysis I’ve segmented it by 30° so there are 12 bands of color. I’ve found that yellow & light greens are the most tricky to pull out accurately from an instinctive perception view of the result. By that I mean what looks very much yellow can get labelled as a bright orange instead as values close to the segment boundary (45°) cause an imbalance towards the orange side.
The basic process is working well for products with one dominating color, for multi-color shots it can pull out the relative levels of each tones hue band but blindly chooses the largest, however slight and uses that for its naming output. A better way would be to do some further analysis on the collected results for each image – one way to enforce some confidence that the main hue really is the main color would be to compare its standard deviation from the average across the sample set. A small σ would indicate a more uniform mix of colors – as we have already sampled out low saturation points this would indicate an image of strong multicolor. Taking the set of hue/tones within the 90th percentile and then comparing their relative deviations of count would allow identifying (and quantifying) the number of colors – so it would be possible to say mainly dark blue with a little light red.
If two bands had similar counts it is also more likley that the color in question is not two separate closely aligned colors but in fact a single color at the midpoint (or thereabouts) of them. A scheme to compare these before the final decision would no doubt improve the detection of yellow which seems to constantly tip into the orange side of the wheel. All in all working on this has been a decidedly pleasant break from data imports, javascript and enough MVC style code to do me a lifetime.
Since I wrote the above I’ve tweaked it somewhat to pull out gray (and black/white) counts also. It works out the rough percentage that corresponds to and if its over a certain limit for black or white will output a value for that also (with a lot of the product shots they’re on a white background so I need to be more granular with that than just regular colors).
#!/usr/bin/python
# -*- coding: utf-8 -*-
# A color analysis script to help you label your store's products with color data
# automagically. It will take either a single file or scour an entire folder for
# folders of images and do each one individually printing a summary of what it
# thinks is the correct color value. A work in progress...
#
# ~jaymz | @jaymzcampbell | jaymz.eu
#
#
# MIT Licesnsed for what its worth, copy: http://www.opensource.org/licenses/mit-license.php
import Image
import ImageFilter
import os
import glob
import sys
import colorsys
import re
from copy import copy
from operator import itemgetter
from decimal import Decimal
output = open('colors.csv', 'w')
# Pixels will be first compared to these values before being
# added to the data list of color information on the first pass
LBOUND = 0
UBOUND = 255
MIN_SATURATION = 30 # avoid washed out pixels influencing counts
# Base folder for the processFolder function, it'll iterate over here on subfolders
FOLDER = '/home/jaymz/documents/crooked-docs/data-export/store-migration/product-images/'
# Meh, i need to flip between these two, you can probably tweak this
SUMMARY_FORMAT, SQL_FORMAT = True, True
# Names based off: http://bluelobsterart.com/wordpress/wp-content/uploads/2009/03/rgb-color-wheel-lg.jpg
COLOR = ['RED', 'ORANGE', 'YELLOW',
'LIME', 'GREEN', 'TURQUOISE',
'CYAN', 'OCEAN', 'BLUE',
'VIOLET', 'MAGENTA', 'RASPBERRY',
]
TONE = ['DARK', '', 'BRIGHT']
# via the createColorSQL.py file , addition added in GRAY/BLACK/WHITE to after this
SQL_IDS = {'DARK YELLOW': 7, 'DARK ORANGE': 4, 'BRIGHT GREEN': 15, 'BRIGHT ORANGE': 6, 'DARK RED': 1, 'BRIGHT OCEAN': 24, 'BRIGHT RED': 3, 'DARK OCEAN': 22, 'YELLOW': 8, 'OCEAN': 23, 'BRIGHT YELLOW': 9, 'RASPBERRY': 35, 'GREEN': 14, 'BRIGHT TURQUOISE': 18, 'CYAN': 20, 'MAGENTA': 32, 'RED': 2, 'ORANGE': 5, 'BLUE': 26, 'TURQUOISE': 17, 'LIME': 11, 'BRIGHT LIME': 12, 'DARK MAGENTA': 31, 'DARK LIME': 10, 'BRIGHT MAGENTA': 33, 'BRIGHT VIOLET': 30, 'DARK VIOLET': 28, 'DARK BLUE': 25, 'BRIGHT BLUE': 27, 'VIOLET': 29, 'BRIGHT RASPBERRY': 36, 'DARK TURQUOISE': 16, 'DARK CYAN': 19, 'BRIGHT CYAN': 21, 'DARK GREEN': 13, 'DARK RASPBERRY': 34}
pcnt = 0
def trimFloat(val, places=2):
return float(repr(val)[0:places+2])
def withinBounds(allowance, _rgb):
rgb = copy(_rgb)
diff = 0
allowance = Decimal(repr(allowance))
for c in rgb:
for d in rgb:
dec_d = Decimal(repr(d)).quantize(allowance)
dec_c = Decimal(repr(c)).quantize(allowance)
diff = abs(dec_d-dec_c)
if (d != c) and diff>allowance:
return False
return True
def processImage(i, name=None):
""" Scales down the image, blurs it to ease the blending of the color values
and reduce spikes from anomolies. It then samples pixels creating a list of
colors. This list is then looped over to build counts which are placed into
bins of 30° hue's seperated into three based on their value. Pixels less than
a certain saturation are discarded. """
global pcnt
i = i.resize((200,200))
i = i.convert("RGB")
i = i.filter(ImageFilter.BLUR)
d = i.getdata()
cnt = 0
h = [] #holds the hsv info
grays = [] #holds just gray content
black_count = 0
white_count = 0
total_samples = 0
for p in d:
cnt = cnt + 1
if cnt == 8: #take every 4th pixel
if p[0]>LBOUND and p[1]>LBOUND and p[2]>LBOUND and p[0]<UBOUND and p[1]<UBOUND and p[2]<UBOUND:
r = trimFloat(float(p[0])/255)
g = trimFloat(float(p[1])/255)
b = trimFloat(float(p[2])/255)
if not withinBounds(0.02, (r,g,b)):
h.append(colorsys.rgb_to_hsv(r,g,b))
else:
if (r+g+b)/3>0.94:
white_count += 1
elif (r+g+b)/3<0.3:
black_count += 1
else:
grays.append(colorsys.rgb_to_hsv(r,g,b))
total_samples += 1
cnt = 0 #reset sample counter
h.sort()
grays.sort()
bin_width = 30 # size of hue slices (degress)
max_bin = 360
darks = [0] * int(max_bin/bin_width)
mids = [0] * int(max_bin/bin_width)
lites = [0] * int(max_bin/bin_width)
for p in h[::]:
hue = p[0]*360
sat = p[1]*100
val = p[2]*100
if sat >= MIN_SATURATION:
bin_number = ((int(hue)+15)/bin_width)%(max_bin/bin_width)
if val<33:
darks[bin_number] += 1
elif val>33 and val < 66:
mids[bin_number] += 1
else:
lites[bin_number] += 1
#print "HUE BIN: %s VALUE : %d" % (int(hue)/bin_width, int(hue))
c = 0
data = zip(darks, mids, lites)
if SUMMARY_FORMAT:
for x in data:
print '%d %s : %s %d°' % (c, COLOR[c], x, c*bin_width)
c += 1
# the following area needs a rework. the index technique works alright as long
# as counts and values dont all match up, then it starts picking the first one
# so this needs re-writing to better order the list data
darks_sort, mids_sort, lites_sort = darks[::], mids[::], lites[::]
darks_sort.sort()
mids_sort.sort()
lites_sort.sort()
sorted_counts = (darks_sort, mids_sort, lites_sort)
primary_idx = (darks.index(sorted_counts[0][-1]), mids.index(sorted_counts[1][-1]), lites.index(sorted_counts[2][-1]))
primary_cnts = (darks[primary_idx[0]], mids[primary_idx[1]], lites[primary_idx[2]])
tone = primary_cnts.index(max(primary_cnts))
max_hbin = primary_idx[tone]
pcnt += 1
if SUMMARY_FORMAT:
print "\nDominant Hue: %s %s" % (TONE[tone], COLOR[max_hbin])
if SQL_FORMAT and name and max(primary_cnts) > 30:
output.write('%d, %s, %s\n' % (pcnt, name, SQL_IDS[' '.join([TONE[tone], COLOR[max_hbin]]).strip()]))
sorted_counts[0][-1], sorted_counts[1][-1], sorted_counts[2][-1] = (0, 0, 0) # kind of reset the primary to null
for l in sorted_counts:
l.sort()
primary_idx = (darks.index(sorted_counts[0][-1]), mids.index(sorted_counts[1][-1]), lites.index(sorted_counts[2][-1]))
primary_cnts = (darks[primary_idx[0]], mids[primary_idx[1]], lites[primary_idx[2]])
tone = primary_cnts.index(max(primary_cnts))
max_hbin = primary_idx[tone]
if SUMMARY_FORMAT:
print "Secondary Hue: %s %s" % (TONE[tone], COLOR[max_hbin])
if SQL_FORMAT and name and max(primary_cnts) > 30:
pcnt += 1
output.write('%d, %s, %s\n' % (pcnt, name, SQL_IDS[' '.join([TONE[tone], COLOR[max_hbin]]).strip()]))
# area to rewrite ends...
gray_total = [(g[0]+g[1]+g[2])/3 for g in grays]
gray_average = reduce(lambda x,y : x+y, gray_total)/len(gray_total)
black_percent = black_count/float(total_samples)*100
gray_percent = len(gray_total)/float(total_samples)*100
white_percent = white_count/float(total_samples)*100
if SUMMARY_FORMAT:
print "\nAverage Gray: %s (samples: %0.1f%%), White count: %s (%0.1f%%), Black count: %s (%0.1f%%)" % (gray_average, gray_percent, white_count, white_percent, black_count, black_percent)
print "Total samples taken: %s\n\n" % total_samples
if SQL_FORMAT:
if black_percent > 10:
pcnt += 1
output.write('%d, %s, %d\n' % (pcnt, name, 38))
if gray_percent > 10:
pcnt += 1
output.write('%d, %s, %d\n' % (pcnt, name, 37))
if white_percent > 30:
pcnt += 1
output.write('%d, %s, %d\n' % (pcnt, name, 39))
# Helper functions follow along with __main__ def
def processFolder(folder):
for image_folder in glob.glob(folder+'*'):
try:
folder_images = []
for image in os.listdir(image_folder):
if "jpg" in image and "._" not in image:
folder_images.append(image)
folder_images.sort()
j = os.path.join(image_folder, folder_images[1])
if SUMMARY_FORMAT:
print "working: "+j
i = Image.open(j)
processImage(i, image_folder.split('/')[-1])
except:
pass
def processFile(_file):
i = Image.open(_file)
processImage(i)
if __name__ == "__main__":
try:
if 'product-images' not in sys.argv[1]:
processFile('product-images/'+sys.argv[1])
else:
processFile(sys.argv[1])
except IndexError:
processFolder(FOLDER)
output.close()