IshiVision

Computer Vision Based System to read the numbers in an Ishihara Plate test

Important note

Under systems like Debian 10 with Cinnamon DE an issues with gtk is present, if that happens to you go to the common.py file and add:
gi.require_version("Gtk", "2.0")
to fix the issue. Any gi/gtk issues can be traced to that missing/extra line depending from your installation.

Used technology:

• Local maxima method to extract the dominant colors per each plate
• Various OCR algorithm such as:
  • kNN
  • SVM
  • GNB
  • Area matching
  • SSD
  • SAD
  • SIFT

Documentation

See here for more informations

Help:

-h, --help Displays this help
-k, --ocr <type> Select the type of ocr
-t, --train Trains the ocr
-s, --size <int> Selects the size of the train set [default = 2]
-l, --load <file> Loads the trained file
-d, --dump <file> Saves the trained data
-v, --verbose Verbose prints
--debug Enables debug features
-a, --accuracy <int> Calculates the accuracy
-c, --char <char> Specify the char to test
--gkt Enables gkt fixes for debian 10 and OpenCV 3.something
--silent Produce no output
-j <json file> Select ocr modules file
-p, --show Show the images and the internal elaboration passages

Available −k parameter

• none - the default test one, will not return a result
• sift - the sift ocr implementation will be run, only -t is supported and -l, -d may not be used.
• knn - the knn ocr will be run, it will require either a data set passed via -l or to generate one via -t.
• svm - the svm ocr will be run, it will require either a data set passed via -l or to generate one via-t.
• sksvm - the sksvm ocr will be run, it will require either a data set passed via -l or to generate onevia -t.
• gnb - the gnb ocr will be run, it will require either a data set passed via -l or to generate one via -t.
• area - the area ocr will be run, only -t is supported and -l, -d may not be used.
• sad - the sad ocr will be run, only -t is supported and -l, -d may not be used.
• ssd - the ssd ocr will be run, only -t is supported and -l, -d may not be used

Examples

Training a kNN on a 10 sized train set and checking the accuracy with30 generated images for two consecutive times
$ python3 test.py -k knn -d data_set -t -s 10 -a 30 --verbose
$ python3 test.py -k knn -l data_set -a 30 --verbose
Run to detect the number 0 in a generated ishihara plate with area matching
$ python3 test.py -k area -t --char 0 --verbose --show
Run to calculate accuracy with 10 images in a generated ishihara plate with area matching
$ python3 test.py -k area -t --accuracy 10 --verbose --show

Generator

It's also possible to generate custom Ishiara plates for testing. To do that run generator.py.

Help

-h, --help Displays this help
-g, --glyph <glyph> Render the given glyph
bg <comma separated hex> Uses the given colors for the background (e.g "0x000000, 0xAAAAAA")
-fg <comma separated hex> Uses the given colors for the foreground
-v, --verbose Verbose prints
<file name> Will be saved as a .PNG regardless

Credits

This program was made for a Computer Vision course in a Master Degree in Computer Engineering by:

• davidegiordano [[email protected]]
• chkrr00k [[email protected]]
  (Code was discussed before actual implementation, don't consider commit numbers as an indication of work done per person)