Automatic Privacy Detection in Scanned Document Images Based on Deep Neural Networks

Abstract

The authors consider the problem of automatic detection of private scanned documents based on text recognition with deep neural networks. The paper suggests implementing a two-phase approach with the first stage which includes efficient EAST text detection and recognition using Tesseract OCR Engine. Secondly, the authors classify the privacy of a scanned document by deep neural networks applied to the extracted text. After that, a special dataset is gathered in order to train these networks. The experiments show that using OCR Engine for both text detection and segmentation ends up with relatively poor identification of private documents when compared to preliminary text detection with EAST method. Moreover, conventional keyword spotting using the list of sensitive words is less accurate when compared to neural network-based methods. Finally, it was demonstrated that the classification of a bag of most frequent words outperforms traditional text classification techniques with LSTM and convolutional networks.