Classification of photographed document images based on deep-learning features

Abstract

In this paper, we propose two new problems related to classification of photographed document images, and based on deep learning methods, present the baseline solutions for these two problems. The first problem is that, for some photographed document images, which book do they belong to? The second one is, for some photographed document images, what is the type of the book they belong to? To address these two problems, we apply “AexNet” to the collected document images. Using the pre-trained “AlexNet” on the ImageNet data set directly, we obtain 92.57% accuracy for the book-name classification and 93.33% accuracy for the book-type one. After fine-tuning on the training set of the photographed document images, the accuracy of the book-name classification increases to 95.54% and that of the booktype one to 95.42%. To our best knowledge, although there exist many image classification algorithm, no previous work has targeted to these two challenging problems. In addition, the experiments demonstrate that deep-learning features outperform features extracted with traditional image descriptors on these two problems.