The project collects the image dataset of Georgian printed letters and recognizes them using CV and CNN from keras and tensorflow. Also there are trainers for learning letters and numerals.
Libraries: cv2, keras, tensorflow, pandas, numpy, re, xml.etree.ElementTree, glob, os, shutil, google.colab, opendatasets, collections, matplotlib, BeautifulSoup4
-
The image dataset of Georgian signboards was created by myself during my living in Georgia. More than 4000 letters on these images were labeled with makesense.ai. The result dataset was uploaded on Kaggle.
-
The data about Georgian alphabet and frequency of letter occurence was scraped from Wikipedia. The georgian_labels.txt file is used for labeling.
-
The first problem is Image Processing. I created a new dataset of cropped images of singular letters in common design (32x32, black letter on the white background). The received dataset (~20MB) was saved as cropped_images.zip archive.
-
The second problem is Image Classification. I built the CNN model based on images of 7 most frequent occurring letters using keras and tensorflow and get the accuracy of 99%. All the samples of some of top-7 letters from external image were classified correctly. The code is in Georgian_letters_recognition_TF.ipynb file.
-
The first trainer is Letter Memorization. It realizes the memorization technique of sequential replacement of letters in Russian or English texts.
-
The second trainer is Numerals Constructor. It generates a Georgian transcription of numerals by any number between 0 and 1 million.