Development of Advanced Techniques for Pattern Recognition

Abstract

Pattern Recognition (PR) is the ocean of paradigms and its applications exist almost in every domain. In this modern computing world, each domain like biometric systems, medical diagnosis, banking, remote sensing, voting, forecasting etc. has huge amount of data in the diverse form. The main objective of PR is to recognize, classify, and analyse such data to make inferences. The PR algorithms fall under the umbrella of machine learning which can be broadly classified in two categories : supervised, and unsupervised algorithms. In supervised learning, the training of model depends only on the labeled data available. In contrast, unsupervised learning paradigms are developed based on the unlabeled data. There exist huge number of supervised and unsupervised learning paradigms in literature and can be classified into different categories. In the present work, the objective is to develop few advanced pattern classification techniques based on both the paradigms which can be applied on various applications of different domains for effective classification. More specifically, the present dissertation accomplishes the following objectives: 1. Developed a semi-supervised learning technique that selects the transductive samples by incorporating new criteria in sample selection process. The proposed technique out performs in two real situations: i) when the initial training samples are biased and, ii) when the initial training samples set is poor. 2. A new fast partition based batch mode active learning technique based on SVM classifier has been developed which gives high accuracy even if the initial SVM is poor. A novel partitioning method has been designed which first divides the unlabeled samples into partitions in one-dimensional feature space according to their distribution in the original feature space. Then to select the most informative samples from the unlabeled pool, one sample from each partition is selected based on an uncertainty criterion defined by exploiting SVM classifier. The number of unlabeled samples selected at each iteration of active learning is determined automatically and depends on the number of non-empty partitions generated. 3. Developed thresholding techniques that mitigate all the limitations of existing traditional thresholding techniques based on histogram of the image. A novel energy curve of the image has been designed which includes the spatial contextual information of image. It has better discriminatory capabilities in comparison to histogram of the image. This energy curve has been utilized to develop thresholding techniques that determine the number of objects present in the image automatically v and gives the optimal threshold values for image segmentation. Extensive experimental work has been carried out to develop number of novel pattern classification algorithms. To assess the effectiveness of the newly designed algorithms results have been compared with the respective state-of-the-art techniques cited in the literature. Experiments have been performed over the number of standard datasets for each technique designed.