http://localhost:8081/jspui/handle/123456789/15073 Wed, 21 Oct 2026 22:23:29 GMT 2026-10-21T22:23:29Z http://localhost:8081/jspui/handle/123456789/21756 Title: SCENE TEXT IMAGE SUPER-RESOLUTION USING DEEP LEARNING Authors: Singh, Shrey Abstract: Scene Text Image Super-Resolution (STISR) is a vital sub-task in document analysis and optical character recognition (OCR), particularly in real-world applications such as license plate recognition, translation, assistive technologies, and augmented reality. In unconstrained environments, text captured in natural scenes often su!ers from motion blur, low resolution, defocus, and other degradations, making it di"cult for both humans and machines to interpret the content accurately. Traditional superresolution techniques, primarily designed for natural images, often fail to preserve the fine-grained semantic and structural features of text, such as strokes, curves, and character boundaries. This thesis addresses these challenges by proposing a suite of di!usion-based generative models that incorporate structural, semantic, and linguistic priors to restore high-quality, readable text images. The first contribution introduces the Di!usion-Conditioned Di!usion Model (DCDM), which integrates low-resolution visual input with character-level semantic embeddings derived from a latent text di!usion module. The second contribution presents the Skeleton-Aware Di!usion Model (SADM), where refined skeleton maps act as structural guides during the di!usion process to improve character-level fidelity. The third and final contribution explores on the stroke level by the Stroke-Conditioned Di!usion Model (SCDM), focusing on low-resource Indic scripts such as Hindi, Marathi, and Gujarati. SCDM leverages a novel stroke decomposition strategy to generalize across linguistically similar scripts through structural supervision. Tue, 01 Apr 2025 00:00:00 GMT http://localhost:8081/jspui/handle/123456789/21756 2025-04-01T00:00:00Z http://localhost:8081/jspui/handle/123456789/20451 Title: A FRAMEWORK FOR METAHEURISTIC BASED ALGORITHMS FOR TEAM FORMATION Authors: Tukaram, Shingade Sandip Abstract: Effective collaboration in networks is crucial for successful team formation. The team forma tion problem involves selecting a subset of agents, referred to as a team, from a larger pool, ensuring the team meets certain desirable properties. This research focuses on selecting agents with the necessary skills, previous communication, and shared abilities, thereby minimizing communication costs. A practical application of the proposed approach is in team formation for IT projects and other team selection scenarios. In this study, we use real-world datasets, ACM,Academia Stack Exchange, DBLP and Players_20 football team dataset to evaluate our methods. Wesuggest a single-objective heuristic approach based on the Grey Wolf Optimizer (GWO) with a modified swap operation to improve upon previous team formation work. This method effectively minimizes communication costs while selecting agents with the required skills. Ex perimental results show that the Improved GWO significantly outperforms traditional methods in terms of both performance metrics and communication cost reduction. Building on this, we propose a hybrid metaheuristic approach that combines Particle Swarm Optimization (PSO) and the Jaya algorithm with a modified swap operator(PSO-Jaya). The third approach focuses on improving algorithm efficiency by integrating state space re duction techniques into the metaheuristic framework to address the increasing complexity and computational demands of the previous methods. The Employee Bee Algorithm (EBA) is en hanced with state space reduction, speeding up the computation while maintaining or improving result quality(IEB). Lastly, we consider a multi-objective optimization context for team formation. For this, we compare several metaheuristic approaches, including NSGA-II, NSGA-II with Simulated Annealing (NSGA-II-SA), NSGA with PSO (NSGA-II-PSO), and our approach Differential Evolution-based NSGA-II (NSGA-II-DE). Sun, 01 Sep 2024 00:00:00 GMT http://localhost:8081/jspui/handle/123456789/20451 2024-09-01T00:00:00Z http://localhost:8081/jspui/handle/123456789/20447 Title: INFORMATION RETRIEVAL AND LOCALIZATION IN DOCUMENT IMAGE Authors: Ali, Tofik Abstract: Information retrieval and localization in document images involve extracting, identi fying, and making accessible the relevant information within digital or digitized visual representations of traditional paper-based documents. This process is crucial for manag ing a vast and diverse range of documents encountered in various sectors such as legal, medical, academic, and corporate environments. Document images, which preserve the content, format, and sometimes the texture of the original documents, play a significant role in maintaining the integrity and authenticity of information. Effective retrieval and localization of information from these images require sophisticated techniques in image processing, machine learning, and deep learning to address challenges such as varying image quality, diverse document formats, and the need for accurate and efficient text recognition and interpretation. The goal is to transform static document images into dynamic, actionable data sources, enhancing their utility and accessibility in real-world applications. The digital transformation has revolutionized how information is stored, accessed, and managed across various sectors. Digitized documents offer numerous advantages over their physical counterparts, including ease of access, improved storage efficiency, and enhanced security. However, the real challenge lies in making this digitized information accessible and intelligible to users. Advanced technologies are required to bridge the gap between digitized information and its practical utility, necessitating the development of robust models and algorithms for efficient processing and interpretation. This research addresses the challenges inherent in document image analysis, such as i Abstract variability in image quality, diverse document formats, and complex layouts. It aims to develop advanced computational models for document image analysis to improve the accuracy and efficiency of character recognition, text segmentation, and image understanding. The study focuses on employing multi-task pre-training strategies to enhance the accuracy and efficiency of these technologies. The research methodology involves breaking down the problem into manageable components and systematically addressing each challenge using convolutional neural networks (CNNs), advanced text segmentation and recognition algorithms, and image understanding techniques. Key contributions of this research include the development of high-accuracy character recognition systems, particularly for handwritten scripts, leveraging advanced CNNs; the introduction of the Gated Multiscale Input Feature Fusion (GMIF) scheme for scale-invariant text detection; the development of Fast&Focused-Net (FFN) for small object feature encoding using the Volume-wise Dot Product (VDP) layer; and the introduction of a multi-task pre-training approach that combines text, image, and layout information to enhance document information analysis. The proposed models and techniques have been evaluated on various datasets, demonstrating significant improvements in the accuracy and efficiency of document image analysis tasks. The real-world applications of these advanced technologies are vast and varied, spanning academic institutions, corporate environments, legal industries, and the medical field. This research contributes to transforming static document images into dynamic, actionable data sources, supporting automated workflows, facilitating decision-making, and promoting knowledge discovery. Keywords: Document Image Analysis, Information Retrieval, Text Localization, Machine Learning, Deep Learning, Convolutional Neural Networks (CNNs), Multi-Task Pre-Training, Image Processing, Text Segmentation, Character Recognition, Gated Multiscale Input Feature Fusion (GMIF), Fast&Focused-Net (FFN), Volume-wise Dot Product (VDP) Layer, Entity Recognition, Relationship Extraction, Layout Analysis. Mon, 01 Jul 2024 00:00:00 GMT http://localhost:8081/jspui/handle/123456789/20447 2024-07-01T00:00:00Z http://localhost:8081/jspui/handle/123456789/20331 Title: SIZE AND PRODUCTIVITY METRICS BASED EFFORT ESTIMATION FOR SOFTWARE DEVELOPMENT Authors: Shukla, Suyash Abstract: Attaining accuracy in effort estimates can yield notable benefits in project planning and is crucial in facilitating efficient project management, frequently exhibiting a substantial correlation with the triumph of a software project. The research community has recently become increasingly interested in applying machine learning (ML) techniques for software effort estimation (SEE). Several researchers have employed different ML techniques for successful SEE over different software datasets. However, selecting the appropriate generalized model for a problem has proven challenging. Obtaining optimal outcomes in SEE by utilizing individual models poses significant challenges. Hence, the researchers proposed an alternate approach wherein multiple models (ensemble) are employed collectively to predict SEE. The ensemble models utilized in previous studies have treated the base learner’s hyperparameter tuning and weight assignment as separate entities, which may result in a trade-off between bias and variance, impacting the ensemble performance. So, we have proposed a SEE model based on the self-adaptive ensembling approach, integrating hyperparameter tuning and model weighting, considering the bias and variance trade-off in decision-making. Also, the SEE datasets contain heterogeneous projects with highly distributed effort values, which may also degrade the prediction model’s performance. So, we have developed a variant of self-adaptive ensembling based on locality to deal with the issues related to data heterogeneity. Accurate SEE is crucial for successfully implementing software projects, and software size plays a major role in it. However, previous methods for estimating effort were founded on metrics such as software lines of code or function points to estimate the size. The increasing need for additional functionalities and the incorporation of new features, such as software reuse, distributed systems, and iterative development, has necessitated the creation of new methodologies for estimating software size and effort. Additionally, previous metrics for software size and approaches for estimating software effort lack automation. They do not utilize Unified Modelling Language (UML) artifacts to reveal software features pertinent to software size. The UML diagrams automatically capture attributes pertinent to the computation of software size. The automation of extracting software size attributes from UML diagrams offers a more efficient approach to calculating software size and estimating software development effort. The Use Case Point (UCP) approach, established by Karner, is a widely recognized and significant early-stage SEE strategy founded on the fundamental elements of use case diagrams (actors and use cases). The utilization of UCP-based approaches is highly appropriate for this particular demand due to its advantageous alignment with two prominent industry practices: (1) the object-oriented (OO) development paradigm and (2) the utilization of use case modeling. Many researchers have conducted investigations utilizing several linear regressionbased models to estimate UCP. While the error estimates derived from existing models demonstrate improvement compared to traditional models, they cannot effectively handle nonlinear interactions within the UCP datasets. So, we have developed UCPbased models to estimate software efforts by utilizing different solo and ensemble models to handle nonlinear relationships in the SEE datasets. Also, the UCP approach consists of size estimation (in UCP) and effort estimation with calculated size. The productivity of a project is one of the main components for estimating effort from the UCP. However, productivity prediction is not explored in the UCP literature. So, we have also proposed a model for productivity prediction based on environmental factors. Mon, 01 Apr 2024 00:00:00 GMT http://localhost:8081/jspui/handle/123456789/20331 2024-04-01T00:00:00Z