Newspaper elements detection and newspaper pages categorization using CNNs and transformers

Newspaper digitization has gained wide interest around the world. Archives of digitized newspapers and magazines contain a wealth of information that spans decades. To extract this abundance of information, optical character recognition (OCR) techniques with extensive manual page annotation have been employed. The OCR techniques extract the text from the raw image of the page, while page annotation adds meta-data about the content of the page such as the category of the page and the location of articles and other elements. To automate this process, I propose a framework for detecting newspaper pages elements and categorizing newspaper pages. The framework will use visual features of the digitized newspaper to classify the printed media type and decompose the newspaper page into its main elements (news articles, advertisements, and page headers) using object detection. Then, it will use both visual and textual features to categorize the newspaper page into its main sections (first page, politics, economy, sports, and advertisement). The element detection stage is leveraged by using only news articles for categorizing the pages, since other elements, such as advertisements, may contain visual and textual features that are not related to the page section. This framework will prepare the newspaper page for the OCR methods to extract meaningful information. The page elements detection phase of the framework is language-agnostic, which allows it to extract the articles from newspapers in different languages (e.g., Arabic, English, French, German, etc). The framework will use two deep neural networks architectures, Faster R-CNN which is based on the convolutional neural network (CNN) architecture and Transformers to classify and detect elements in the printed media.