An Autoencoder-Based Approach of Automatic Writer Identification from Biscriptual Characters

In the last decade, automatic writer identification using a convolutional neural network (CNN) has been well studied. For further performance improvement of the writer identification task, a generative adversarial network-based approach is proposed in this paper to find out the corresponding writer of the given handwriting. This is done using an autoencoder model as the generator and a CNN - based classifier model as the discriminator. The autoencoder network generates synthetic images from the input real images which are then fed to the discriminator along with input real images. The synthetic images provide crucial features to the classifier network. Thus, the autoencoder is used as a feature extractor to aid the discriminator for classification. In this work, autoencoder and discriminator models are trained simultaneously on a self-captured dataset containing handwritten characters from two scripts: Bangla and English written by 24 writers three times. Writer identification accuracy of the proposed autoencoder-based approach is 94.24% with a relative improvement of 2.54 % over the accuracy of 91.85% obtained using the same discriminator without the autoencoder. The proposed classifier also outperforms prior models where the accuracy of the second-best model is 90.07%.