Document recognition method based on convolutional neural network invariant to 180 degree rotation angle

In this work we deal with the problem of recognition of printed document, captured by scanned devices and mobile phones. Recognition of documents' images rotated by 180 degrees, by known approaches involves orientation detection of image, then rotation if necessary, and the actual document image recognition in the correct orientation. The proposed approach based on convolutional neural network that is invariant to the angle of rotation by 180 degrees, eliminates the steps of orientation detection and image rotation. This speeds up the recognition process on mobile platforms, which performance is currently concedes to server and desktop platforms. Recognition of two data sets was considered: scanned images of structured national documents and public SmartDoc dataset, which contains images captured by mobile phones. For this datasets the accuracy of document recognition was estimated. The accuracy of the orientation detection using the proposed method on the considered stands is 100%, which exceeds the accuracy of the orientation detections of the methods described in the works from the list of references.