Abstract
Identifying scripts in natural images is an important step in document analysis. Recently, Convolutional Neural Network (CNN) has achieved great success in image classification tasks, due to its strong capacity and invariance to translation and distortions. A problem with training a new CNN is that it requires a large amount of labelled images and extensive computation resources. Transfer learning from pre-trained models proves to ease the application of CNN and even boost the performance in some circumstances. In this paper, we use transfer learning and fine-tuning in document analysis. Indeed, we deal with the scene script identification quantitatively by comparing the performances of transfer learning and learning from scratch. We evaluate two CNN architectures trained on natural images: AlexNet and VGG-16. Experimental results on several benchmark datasets namely, SIW-13, MLe2e and CVSI2015, demonstrate that our approach outperforms previous approaches and full training.