Klappentext

Documents have been a rich source of the day-to-day medium of communication, and also as a historical and legal record ever since their inception. Due to the advancements in computer technology, birth of scanners and advanced printers, there was a strong motivation to move towards automation and eventually towards a paperless office. Because of which more and more documents started generating in the digital form, which led to the inventions like OCR technology and later many computer assisted document image analysis techniques like enhancement, segmentation, retrieval, indexing, word spotting were developed. However, in the present era of Big data where large volumes of documents are being generated on a daily basis, the documents are generally subjected to compression for the efficiency of storage and communication. Further in order to analyse or process such digitally compressed documents, the conventional way to do this is to decompress, operate, and subsequently recompress it, which indent additional computing resources. Thus, it would be novel and efficient to operate directly over the compressed documents without involving decompression, which is termed as Compressed Domain Processing. Therefore the present thesis is focused on developing novel algorithms to accomplish Document Image Analysis (DIA) directly in the Compressed Domain. Document segmentation is a systematic process where it extracts the layout of the document, separates text and non-text components from printed and handwritten documents. Text segmentation techniques involve separating paragraphs, text-lines, words, and characters for OCR applications or direct recognition / analysis. Most of the existing state-of-the-art document segmentation methods are handcrafted and can work only with uncompressed document images. There are also some recent document segmentation techniques based on the principle of deep learning, but they are designed to work with uncompressed document images. In the literature, there are many compression algorithms and file formats that support document image compression, such as CCITT, JPEG, JPEG2000, PNG, PDF, and BMP. JPEG is the most popular compression algorithm that is widely supported over internet and hardware devices. Since JPEG is the default compression algorithm in mobile phones, digital tablets, digital cameras, scanners etc., more than 90% of the data in the internet world and digital libraries exists in the JPEG compressed form. This gives us strong motivation to propose novel segmentation algorithms for compressed documents that are largely stored in JPEG compressed format. Therefore, the present thesis aims at exploring the document segmentation problem directly in the JPEG compressed domain using both handcrafted and deep learning-based methods. In the backdrop of the above discussions, the research work in this thesis converges to the objective of investigating the possibility of developing document segmentation methods like layout segmentation, text line segmentation and word segmentation directly in JPEG compressed documents. Further, due to the recent successful footprint of deep learning-based models in solving many real time applications, a strong inspiration is developed for exploring deep learning-based document segmentation methods in