DOCUMENT IMAGE ANALYSIS AND RETRIEVAL SYSTEM

Abstract

The drastic development in information technology has a lead for the digitization of documents in each and every field. During the digitization process, many of the existing and newly generated documents will be stored in the form of images known as document images. This has created an opportunity for the researchers to develop techniques and algorithms for analysis of document images for implementation of expert systems. Huge database of document images also require the techniques for accessing, searching and browsing of the documents based on certain criteria. The main objective of the proposed research work is to develop new techniques and algorithms for document image analysis and retrieval. The proposed research work is motivated by the Google search engine which allows text-based searching of information from the World Wide Web. The thesis provides new algorithms, techniques and feature extraction schemes for implementation of document image analysis and retrieval system. Five document retrieval techniques based on the logo, signature, face/photo, fingerprint and language are addressed in the thesis. The thesis first provides an introduction to the document image analysis and retrieval system, its applications and a detailed literature survey. The literature survey discusses about the current state of the art and research trends. An efficient automatic logo-based retrieval technique is proposed by using mathematical tool Singular Value Decomposition (SVD). New feature extraction techniques based on singular value decomposition are used for logobased document image retrieval. The proposed method is tested on the document images of Tobacco 800 database [41]. The experimental results are better compared to the earlier approach [47]. The thesis also provides signature-based document image retrieval method using multi-level Discrete Wavelet Transform (DWT). To investigate the influence of similarity metrics on retrieval performance, experiments are carried out using seven distance metrics namely Euclidean, Canberra, City-block, Chebychev, Cosine, Hamming and Jaccard. The city block distance provided a highest precision of 80% using multi-level DWT features. Recently many documents such as identity cards, passports, driving license etc. are embedded with face/photo of a person. For retrieval of such documents, the thesis proposes face/photo based document image retrieval using Gray Level Cooccurence Matrix based features. In this approach, the cooccurence matrices for Red, Green and Blue components of the face image vii are computed and their diagonal elements are used to construct the feature vector. This helps in reducing the number of features and computational complexity. Experiments are carried out on a database of 810 document images. The proposed method provided a mean average precision of 82.66%. To provide high security fingerprint impression is being used in many important documents such as property registration, banking transactions, insurance related documents etc. This motivated us to propose a fingerprint-based document image retrieval technique using multiresolution Local Binary Pattern (LBP) features in the thesis. The proposed method is tested on 1200 document images having fingerprint impression. A mean average precision of 73.08% is obtained for retrieval of top 1, top 5, top 8, top 15 and top 20 document images. The experimental results are encouraging compared to the existing feature extraction techniques [38] and [39]. This thesis also presents a new approach for classification and retrieval of document images based on the language. The multi-resolution Histogram of Oriented Gradients (HOG) features are proposed for the implementation. The system is tested for printed document images of Kannada, Marathi, Telugu, Hindi and English. Proposed system provided classification accuracy of 87.02% for 1006 document images. The retrieval performance obtained using proposed feature extraction scheme is very promising and encouraging. The thesis also provides future directions for the research to elevate current state-of-the-art.