Text Document Orientation Detection Using Convolutional Neural Networks

Identifying the orientation of scanned text documents has been a key problem in today’s world where every department of any cooperation is surrounded by documents in one or another way. In this paper, our emphasis is on the more challenging task of identifying and correcting the disorientation of general text documents back to normal orientation. Our work aims to solve the real-world problem of orientation detection of documents in PDF forms which can be later used in further document processing techniques. All further document processing tasks depend on detecting the correct orientation of the document. To do this, the convolutional neural network (CNN) is used which can learn salient features to predict the standard orientation of the images. Rather than the earlier research works which act mostly between the horizontal and vertical orientation of non-text documents only, our model is more robust and explainable as it works at page level with text documents. Also, we have accelerated to a different level with proper explanation and interpretability. The proposed approach runs progressively in real time and, in this manner, can be applied to various organizations as well.