Authors: Yash Patel, Michal Bušta, Lukáš Neumann, Jiri Matas
Description: Our method uses a CNN based approach for script-identification in cropped work images. We employ the use of convolutional layers from VGG-16 architecture along with a Global-Average-Pooling and two fully connected layers. Objective of our method is to preserve the aspect ratio of input images. Thus, for both training and testing we resize the images into fixed-height (64) and variable-width ((image width*64)/image height) tensors. For training, we initialize the convolutional layers with ImageNet weights. We make use of categorical-cross-entropy loss function and update all the layers (both convolutional and fully connected) during back-propagation.