DEVELOPMENT OF FRAMEWORK FOR CLASSIFICATION AND ARCHIVING HISTORICAL MANUSCRIPT IMAGES

Abstract

Historical manuscripts are valuable resources for historical information about the distant past. From them culture, education, and ways of life in the past can be gleaned. Due to advancement of Information Communication and Technology (ICT), most of the historical manuscripts have been digitized via scanning devices to electronic formats like digital images. This has resulted in large amounts of historical manuscripts available to public as digital images and other electronic forms. Historical manuscript images (HMI) are easier to manage (by sharing, handling, storage, and processing) compared to actual manuscript documents. In addition, this helps to preserve actual historical manuscripts since they are seldom needed physically. Due to the proliferation of large amounts of HMI, their management and processing is the main focus. HMI management involves a range of tasks and processes carried out on HMI like curation, provenance, indexing and archiving, storage, restoration, retrieval, and classi fication among others. This thesis focuses on development of computer-based framework to index, archive, and classify HMI. First of all, a number of pre-processing tasks are carried out to enhance visual quality of HMI and in turn increase output performance. The pre processing tasks carried out include denoising, binarization, word segmentation and word image normalization. Due to degradations in most of HMI, a model-based binarization technique is proposed to enhance them. In this technique, HMI pixels are modelled to foreground and background pixels by training multilayer perceptron (MLP) classifier using various handcrafted features with high discriminating powers. The features are extracted from HMI. A component tracing and association (CTA) technique has been developed for efficient word vi segmentation of HMI. The merit of the method is in segmenting overlapping and crossing words. Using the concept that short sections of a continuous stroke joined at a common point are symmetric or near symmetric about the common joining point, crossing strokes are identified and separated using (MD − DTWD) multi-dimensional dynamic time warping with dependence. method. A segmentation-based handwritten word spotting (HWS) technique has been developed for indexing HMI. Integral histogram of oriented displacement (IHOD) feature descriptor is used to develop MLP-based HWS system. IHOD descriptor is obtained by computing displace ments of foreground pixels w.r.t centers of their respective m × m cells where m = 15 pixels. Cells are obtained by sub-dividing entire HMI. A fragmented long short-term memory (Frag-LSTM) method is proposed for language iden tification (LID) of textual content of HMI. 3 LSTM networks are used to learn and extract local and global features from input text word. A combined feature vector is obtained by concatenating global and local features. This combined feature vector is then used for LID. Bi-directional fragment network (BiD-FragNet) is proposed for prediction of era or pro duction time of HMI. BiD-FragNet consists of 2 convolution neural network (CNN)-based channels; main and fragment channels. The main channel learns and extracts global fea tures by processing full patches of HMI. Fragment channel is used to learn and extract local features by processing fragments (sub-patches) of HMI. Both channels share information in both directions at various levels. Global and local features learnt are then concatenated to one feature vector which is used with classification layer to give final classification output. Classification output is obtained by voting and averaging schemes. Funnelling ensemble method for writer identification (FEM-WI) has been proposed for HMI. It is a 2-level system of classifier ensembles. in this system, first, 5 newly proposed features (also called base features) are extracted from segmented handwritten words. In level 1, each feature is used to train individual base classifier (MLP). Meta features are then obtained as outputs of level 1 (base) classifiers via k-fold cross validation (KFCV) method. A single level 2 meta classifier that gives final output is trained using the meta features. FEM-WI works by leveraging on different base features for same input word image funnelled to a common vii feature space in level 2 classifier. Thus, writer identification of query word using any one of the base features benefits from all other features used to train meta classifier, hence giving improved output performance.