WriterIdentificationNetwork

Writer Identification Network (WIN)

WIN structure

The header of WIN composes of 4 groups of max-pooling and convolutional layers which produces a local feature vector for each image. Next, n local features of n-tuple images are aggregated and fed into the final softmax layer for identifying the writer.

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

For details of all arguments, please look at this page