http://localhost:8081/jspui/handle/123456789/15072 2026-10-26T12:23:38Z 2026-10-26T12:23:38Z Singh, Shrey http://localhost:8081/jspui/handle/123456789/21756 2026-10-21T11:19:09Z 2025-04-01T00:00:00Z

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.

2025-04-01T00:00:00Z Bagde, Abhishek http://localhost:8081/jspui/handle/123456789/21597 2026-09-20T07:19:05Z 2023-06-01T00:00:00Z

Title: Enhancing Few-shot Semantic Segmentation Technique for Medical Image Segmentation Authors: Bagde, Abhishek Abstract: When confronted with new, unseen semantic classes, fully-supervised deep learning seg mentation models suffer, as their fine-tuning necessitates enormous quantities of data with annotation. Few-shot semantic segmentation (FSS) offers a solution by enabling the seg mentation of previously unseen semantic classes by using only a few labelled examples and without any reliance on fine-tuning. Existing state-of-the-art FSS approaches primar ily focus on segmenting natural imageset and rely on abundant annotated training data for effective generalization to testing classes that are unseen. Due to the above reasons, this training technique is not practical in clinical imaging situations with scarce annotations. To overcome the difficulties mentioned above, our work proposes enhancement in the exist ing self-supervised FSS framework for medical images called SSL-ALPNet; this framework eliminates the need for annotations which were required in the training phase and leverages superpixel-based pseudo-labels that provide supervision signals and also includes an adap tive local prototype pooling module integrated into the prototypical networks for improvis ing segmentation accuracy. The efficacy ofourmodelisdemonstratedinorgansegmentation using Abdominal-MRI images. Experimental results show that our model, which is an en hancement made to the SSL-ALPNet [1] approach, achieves better segmentation accuracy as compared to other standard Few-Shot-Segmentation methods whose training is based on a manually annotated dataset.

2023-06-01T00:00:00Z Gujar, Ameya http://localhost:8081/jspui/handle/123456789/21596 2026-09-20T07:18:49Z 2023-06-01T00:00:00Z

Title: Hybrid Model for Impact Analysis of Climate Change on Droughts in Maharashtra State Authors: Gujar, Ameya Abstract: Droughts are prolonged periods of dry weather that have become more frequent and severe due to climate change and global warming. It can have devastating effects on agriculture, water resources, and ecosystems. Hence, a framework for the prediction of droughts is necessary for mitigating its impact, as it enables authorities to prepare and respond effec tively. This paper presents a hybrid model comprised of the Convolutional Neural Network and Gated Recurrent Units (called CNN-GRU) to predict the Standardized Precipitation Evapotranspiration Index (SPEI), which is used to measure drought intensity. We used India Meteorological Department (IMD) rainfall and temperature data of Maharashtra state of India during the years 1960-2021 as the historical dataset. Whereas, for the future pro jections, we used the Coupled Model Intercomparison Project Phase 6 (CMIP6) dataset of the same region during the years 2015-2100 for different Shared Socioeconomic Pathways (SSP) scenarios. Both these datasets include the daily precipitation, minimum temperature and maximum temperature values. The proposed model is trained and validated using IMD dataset and the final evaluation of its ability to predict the future droughts is conducted on the CMIP6 dataset. We confirmed that it outperforms in terms of mean squared error, mean absolute error, and root mean squared error over both the IMD and the CMIP6 datasets based on the comparative study with the existing deep learning models.

2023-06-01T00:00:00Z Bhardwaj, Ankit http://localhost:8081/jspui/handle/123456789/21595 2026-09-20T07:18:34Z 2023-06-01T00:00:00Z

Title: Impact Analysis of Data Augmentation on Neural Network Models Authors: Bhardwaj, Ankit Abstract: Data augmentation is a widely employed technique in deep learning to expand the training dataset artificially[1]. Its effectiveness in enhancing neural network models’ performance across computer vision and natural language processing tasks has been widely demonstrated[2]. This thesis presents a comprehensive study on the impact of various data augmentation techniques on neural network models. Experimentswereconductedonimagedataset, evaluatingdifferentneu ral network architectures with and without dataaugmentation. Theresultsreveal significant improvements in model accuracy and robustness due to dataaugmen tation. However, the choice of augmentation technique and parameters is crucial. experiments on YOLO[3] and Faster RCNN[4] object detection models demon strate notable performance enhancements with data augmentation. These find ings underscore the efficacy of data augmentation in improving deep learning model performance, particularly for object detection tasks.

2023-06-01T00:00:00Z