deveth0/opencv-characterclassification repository overview

⁠Character Classification with a synthetic data set

This sample project creates a dataset of characters (A-Z, 0-9) with some random parameters (like font color or blur) and uses it to train two different models (KNearest and Support Vector Machine/SVM).

To validate the accuracy of the models, 10% of the generated data is used for testing and 90% for training. After running the training, you'll see the two test datasets (one for KNearest, one for SVM) and all characters that where not classified correctly are marked in red.

For further usage the SVM model is saved in target/chars_svm.dat and can be used in your own code:

import cv2 as cv
import os
import sys

saved_model = "target/chars_svm.dat"
if not os.path.exists(saved_model):
    sys.exit('"%s" not found, run train.py first' % saved_model)

model = cv.ml.SVM_load(saved_model)

⁠Data Generation

First of all you need to generate a data set and use it for training. This is done by the train.py script. If you pass a -s parameter, this defines the number of samples that is created for each character in training (e.g. if you use -s 100 you'll get 100 images of A, 100 images of B etc).

As the generation takes quite a long time, you can skip it for later runs if you want by obmitting the -s parameter.

Example:

# Generate 36*100 images and train the model
python3 train.py -s 100 
# Reuse dataset and train the model
python3 train.py 

The image creation requires ImageMagick 7 to be installed. If you don't want or can't install it on your system, you can also use a Docker image for training (see Docker⁠).

All sources for the image creation can be found in gen_data where you can also add new backgrounds and fonts (or replace the existing ones):

• *.jpg: jpg files are used to create crops (32px x 32px) on which the characters are placed
• *.ttf: ttf files define the font(s) that are used to create characters

⁠Docker

If you have a running Docker installation, you can also use the Docker Image deveth0/opencv-characterdetection⁠. It includes all the background images and fonts from this repository.

To run the image, you need to create a target/ folder and then run the image with a bind mount. When the training is complete, you can find both the created dataset and the saved SVM model in target/.

# Run docker image with training size of 1000
mkdir target
docker run --mount type=bind,source="$(pwd)"/target,target=/target deveth0/opencv-characterclassification "-s 1000"

To replace the gen_data folder in the image, you can mount a local one into the container:

# Run docker image with training size of 1000
mkdir target
docker run --mount type=bind,source="$(pwd)"/target,target=/target --mount type=bind,source="$(pwd)"/gen_data,target=/gen_data deveth0/opencv-characterclassification "-s 1000"

⁠Acknowledgments

The idea to create a syntetic data set originates from Github: learnopencv - CharClassification⁠ and is used in a modified way in this project.

⁠Literature

If you want to find out more about how a Support Vector Machine works, I highly recommend this article⁠ by Savan Patel⁠.