PRL2018_WIN

Overview

This repository contains source code of the paper "Text-Independent Writer Identification using Convolutional Neural Network". Read paper here.

If you want to refer to our paper, please use citation as follows:

Hung Tuan Nguyen, Cuong Tuan Nguyen, Takeya Ino, Bipin Indurkhya, Masaki Nakagawa, Text-independent writer identification using convolutional neural network, Pattern Recognition Letters, 2018, ISSN 0167-8655, https://doi.org/10.1016/j.patrec.2018.07.022.

For questions or more details, please contact us via email addresses: ntuanhung@gmail.com or nakagawa (at) cc.tuat.ac.jp

Prerequisites

Install the following packages via pip3: numpy, tqdm, Pillow, pickle, gzip, scipy, tensorflow-gpu

We developed and evaluated our source code on Tensorflow 1.7.0 with Python 3.6.5.

Notes

Inside images folder, all image file should be named by <writer_id>_<image_name> such as 0_1234. All images should be resized to squared and unique size, for example, 64-by-64.

In configs folder, there are different configuration files which are used to train/evaluate model. Each configuration has three text files for train/valid/test sets.

train-files_<number_of_writers>users_<number_of_patterns_per_writer>patPerUser_SAME_<database_name>.txt
valid-files_<number_of_writers>users_<number_of_patterns_per_writer>patPerUser_SAME_<database_name>.txt
 test-files_<number_of_writers>users_<number_of_patterns_per_writer>patPerUser_SAME_<database_name>.txt

Each configuration file has <number_of_writers> lines. For every line, the first value is <writer_id> which is followed by <number_of_patterns_per_writer> values represented for <image_name>.

Train model

Run the following command to train network

python win5_subimage_classification.py --num_writers 10 --num_training_patterns 100 --agg_mode average --n_tuple 10

If you want to train using GPU, append the following argument

--gpu <gpu_id>

Evaluate model

Run the following command to evaluate the trained model in models folder For example, the evaluate model contains these following files:

  win5_subimg_average@2017.09.22-01.02.03_bestAcc.ckpt-100.data-00000-of-00001
  win5_subimg_average@2017.09.22-01.02.03_bestAcc.ckpt-100.meta
  win5_subimg_average@2017.09.22-01.02.03_bestAcc.ckpt-100.index

<model_filename> should be win5_subimg_average@2017.09.22-01.02.03 and <global_step_eval> is 100.

python win5_subimage_classification.py --training False --model_name <model_filename> --global_step_eval 100 --num_writers 10 --num_training_patterns 100 --agg_mode average --n_tuple 10

Description of arguments

Argument	Type	Description
Fix arguments
--author_name	str	author name
--directory_path, -dp	str	path to PRL2018_WIN
--dataset_name, -dn	str	dataset name
--img_size, -im	int	image size
--image_type, -it	str	image type includes BIN, RGB
--selection_mode, -sm	str	selection mode such as SAME or DIFF
Configuring arguments
--num_writers, -nw	int	number of writers
--num_training_patterns, -ntp	int	number of training patterns per writer
--local_feature, -lf	str	type of local feature type: subimg
--agg_mode, -am	str	different aggregation modes: average, max, kmax
--kmax, -k	int	value of k in kmax aggregation mode
--n_tuple, -nt	int	tuple size
--train_num_permutations, -trnp	int	number of permutations per epoch during training
--valid_num_permutations, -vanp	int	number of permutations per epoch during validating
--test_num_permutations, -tenp	int	number of permutations per epoch during testing
--eval_num_permutations', -evnp	int	number of permutations per epoch during evaluating
Training arguments
--lr, -l	float	learning rate
--max_epochs, -me	int	maximum number of epochs for training
--max_no_best, -mnb	int	early stop if accuracy does not increase during <max_no_best> epochs
--writer_per_batch, -wpb	int	number of writer per minibatch
--gpu, -g	int	gpu_id is used
--gpu_mem_ratio, -gmr	float	memory ratio of gpu from 0.0 to 1.0
Evaluation/Training arguments
--training, -t	bool	if evaluate model using False, else using default (True) for training
--resume, -r	str	if resume training process, pass the model filename
--global_step_start, -gss	int	if resume training process, pass the model global_step to continue training
--model_name, -mn	str	if evaluate model, pass the model filename.
--global_step_eval, -gse	int	if evaluate model, pass the model global_step to evaluate
--use_valid_data, -uvd	bool	use valid data or not while training model
--use_test_data, -utd	bool	use test data or not while training model
--eval_test_data, -etd	bool	use test data or not while evaluating model