/imgsim

backend for cryptodraw chrome extension

Primary LanguagePython

imglib: an image similarity algorithm compendium

A library of fast(ish) image similarity algorithms!

Source of images: favorites gallery of prominent artist on dA

  • chosen because she has similar visual taste--would be representative of most styles that we are likely to see in people using this extension

Original code & work (guide to directory layout)

  • src/fmiq
    • fmiq.py: implements the fast multiresolution querying algorithm (see paper/ directory)
    • fmiq-process.py: parses the output of fmiq on the source images and queries with test images
  • src/opencv
    • histogram.cpp: wrapper around four opencv measures, as well as their queries
      • how to build: run make
    • CMakeLists.txt (getting this to build the right Makefile was not easy)
  • src/phash
    • phash.c: wrapper around phash library (located in lib). generates perceptual hashes of everything in source image directory
      • compatible with radial hashing, DCT hashing
      • how to build: g++ -c phash.c ; g++ -o phash phash.c main.o => generates executable phash
    • phash-process.c: processes the output of phash.c; scores all test images and assigns relative distances
      • how to build: g++ -c phash-process.c ; g++ -o process phash-process.c main.o => generates executable process
  • src/analytics
    • process.py all purpose tool for evaluating the accuracy of the algorithms above
  • data/
    • data/opencv
      • all: results of algorithm on 10 different test data sets
      • opencvresults: output of process.py on those files
      • opencv.md: some data consolidation analysis
    • data/phash
      • radial/: results of radial hashing algorithm on 10 different test data sets
      • dct/: results of dct hashing algorithm on 10 different test data sets
      • radishresults: output of process.py on radial/
      • dctresults: results of process.py on dct/
      • phash.md: some data consolidation and analysis
    • data/fmiq
      • all: results of algorithm on 10 different test data sets
      • fmiqresults: output of process.py on those files
      • fimq.md: some data consolidation and analysis
  • tests/: directories of test data set after various small digital modifications
    • all folders here created through image manipulation by imagemagickcommand line tool

Requirements & dependencies (to be expanded)

  • openCV
  • Boost

Log

Thursday (worked 4.5 hours)
  • gathered source images
  • narrowed down similarity algorithms
  • implemented part of multiresolution querying
  • set up mySQL database for front-end
Friday (worked 8 hours)
  • set up pHash (perceptual hash)
    • installed MacPorts, libpng, CImg dependencies
    • ...before downloading source and copying it into the directory
    • wrote wrapper program phash.c
    • testing two hashes: radial and regular
  • set up openCV histogram comparison method (a library for C++)
    • set up cMake
    • installed Boost and related libraries for iterating through directories
    • learned basic C++
    • wrote wrapper program histogram.cpp
  • finished implementing multiresolution querying (see fmiq)
    • fixed PIL pip install
    • attempted optimizations using pypy (didn't work)
    • each image takes about 7 seconds to analyze
    • straight from paper; rgb to yiq color space conversion
  • homogenized source images
    • installed imagemagick through macports
    • converted all files to jpg
    • converted all jpg to rgb color space
    • set up test files, including effects
Saturday (worked 7 hours)
  • set up test images
    • 20 for each effect (see tests folder)
    • aspect ratio: fixed aspect ratio (300x300px)
    • contrast: increased histogram contrast
    • crop: cropped images to squares, up to 1/2 of image cropped off
    • grayscale: converted to grayscale
    • hue: modulate histogram
    • paint: blobby effect
    • resizedown, resizeup: 50% and 200%, respectively
    • rotate: rotate 90 degrees
    • tintred: duochrome
    • txt: todo, overlaid text
  • tested for duplicates among source images
  • fixing fimq: tried for 30 minutes to install PIl, gave up and installed a virtualenv instead
  • completed fimq implementation
  • completed fimq processing and porting
Sunday (worked 5 hours)
  • debugged fimq porting
    • this took a ridiculously long time.
    • started data collection!!!
  • set up wrapper files for phash: still working on radial though
    • solution to exporting signature data: write to file in a specific format
  • now both phash and fimq are ready for data collection
Monday (worked 9 hours)
  • get Boost working and compile the thing
  • debug the program
  • wrote a program to process output
  • analyzed and wrote evaluations for fmiq and opencv
  • collected data for fmiq and opencv
  • analyzed data for fmiq and opencv
  • wrote most of presentation
Tuesday (worked 4 hours)
  • fixed pHash
  • conducted pHash evaluations
Wednesday (worked 3.5 hours)
  • implemented radial hashing
  • evaluated radial hashing results
  • cleaned up and documented code