Visual unsupervised deep learning model design for historical document image analysis

Abstract

Historical documents are one of the most crucial influences that drive scientific and historical development. Some historical documents are present and can be used through classical models to be analyzed. Other documents do not meet the quality and minimum visibility required by information retrieval systems. Furthermore, the ancient manuscript analysis models include various algorithms and techniques to make document images more understandable for computers. Although classical techniques have mostly overcome the issue of analyzing and extracting information from such documents, the task of visual information, including enhancement and segmentation, is still a demanding task. Due to the complex characteristics of historical documents and their nature of degradation, document image processing has always been an essential task. Many existing approaches, including text and ornament recognition, achieve the information by measuring the width, height, and aspect ratio. Since most historical documents are handwritten, such approaches fail to analyze such sensitive data. Besides, the issues at the technical levels are the enhancements because of poor segmentation results at the noisy level of historical document images. Exploring and pursuing the actual visual objects that enhance the entire ancient manuscript would help us convey a more reliable historical document representation. These visual objects can be tables, figures, characters, ornaments, shapes, and also the entire page. This thesis concerns the design of machine learning tools for more accurate detection of various objects on historical documents and establishing a framework for each of the driven objectives. The proposed approaches promote the usage of deep learning models for compactly enhancing the quality of data. In particular, we will argue how to learn from mapping colour data onto binary (noise-free) in order to remove the degradation. Then, we will describe an unsupervised approach for simultaneous object segmentation in an unsupervised manner. Have it all over, in this thesis, we focus on two such techniques, namely historical document image enhancement, where we will highlight an inference of generative adversarial networks for extracting pixels from an image in order to produce the final binary document image result with better quality. In the present study, we propose an effective deep convolutional generative adversarial network with a few additional parameters that can be trained on various document images to manage the complexity of historical documents and remove degradation. Furthermore, the deep segmentation network can accurately segment the visual objects of historical documents through mapping their data points in the different clusters. The generalization capability and robustness of the proposed framework can remove degradations and segment pages containing characters and ornament regardless of their texture and layouts. This depiction enhances upon document binarization and provides more actual estimation. Experimental results are shown on numerous databases, including READ-BAD, c-BAD, IAMHist, DSSE and DIVA-HistDB.We also present the results of our two articles which are published in ICPR 2020.

Date	6 Jun 2022
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Mohamed Cheriet (Supervisor)