A neural network based detector for handwritten words.

• Download trained model, and place the unzipped files into the model directory
• Go to the src directory and execute python infer.py
• This opens a window showing the words detected in the test images (located in data/test)
• Required libs: torch, numpy, sklearn, cv2, path, matplotlib

• The model is trained with the IAM dataset
• Download the forms and the xml files
• Create a dataset directory on your disk with two subdirectories: gt and img
• Put all form images into the img directory
• Put all xml files into the gt directory

• Go to src and execute python train.py with the following parameters specified (only the first one is required):
  • --data_dir: dataset directory containing a gt and an img directory
  • --batch_size: 27 images per batch are possible on a 8GB GPU
  • --caching: cache the dataset to avoid loading and decoding the png images, cache file is stored in the dataset directory
  • --pretrained: initialize with saved model weights
  • --val_freq: speed up training by only validating each n-th epoch
  • --early_stopping: stop training after n validation steps without improvement
• The model weights are saved every time the f1 score on the validation set increases
• A log is written into the log directory, which can be opened with tensorboard
• Executing python eval.py evaluates the trained model

• The model classifies each pixel into one of three classes (see plot below):
  • Inner part of a word (plot: red)
  • Outer part of a word (plot: green)
  • Background (plot: blue)
• An axis-aligned bounding box is predicted for each inner-word pixel
• DBSCAN clusters the predicted bounding boxes
• The backbone of the neural network is based on the ResNet18 model (taken from torchvision, with modifications)
• The model is inspired by the ideas of Zhou and Axler
• See this article for more details