A simple CNN for n-class classification of document images.
It doesn't take colour into account (it transforms to grayscale). For small numbers of classes (2 to 4) this model can achieve > 90% accuracy with as little as 10 to 30 training images per class. Training data can be provided in any image format supported by PIL.
pip install document-classifier
or
poetry add document-classifier
from document_classifier import CNN
# Create a classification model for 3 document classes.
classifier = CNN(class_number=3)
# Train the model based on images stored on the file system.
training_metrics = classifier.train(
batch_size=8,
epochs=40,
train_data_path="./train_data",
test_data_path="./test_data"
)
# "./train_data" and "./test_data" have to contain a subfolder for
# each document class, e.g. "./train_data/letter" or "./train_data/report".
# View training metrics like the validation accuracy on the test data.
print(training_metrics.history["val_acc"])
# Save the trained model to the file system.
classifier.save(model_path="./my_model")
# Load the model from the file system.
classifier = CNN.load(model_path="./my_model")
# Predict the class of some document image stored in the file system.
prediction = classifier.predict(image="./my_image.jpg")
# The image parameter also taks binary image data as a bytes object.The prediction result is a 2-tuple containing the document class label as a string and the confidence score as a float.
- Give every CNN instance its own isolated tensorflow graph and session
- Fix a bug that occured when using multiple model instances at the same time
The model architecture is fixed for now and geared towards smaller numbers of classes and training images. I'm working on automatic scaling for the CNN.