Air Hockey: An Android-Based Game is a simple 2D arcade game developed using Unity3D to explore the techniques and concepts involved in game development. The game features both local multiplayer and single-player modes, where players attempt to score goals by directing a puck across a frictionless board using a controllable handle. In single-player mode, users compete against an AI-controlled opponent, providing an engaging and fast-paced gaming experience. The project aims to familiarize developers with core game development principles while offering users a quick and entertaining game suitable for short breaks. System and user experience testing demonstrated satisfactory performance and gameplay quality.

This project focuses on developing a News Sentiment Classification System that predicts whether a news article carries a positive or negative sentiment. Motivated by research suggesting that negative news can adversely affect readers’ mental well-being, the system aims to help users identify the nature of news before reading it. The project utilizes web crawling techniques to collect news articles from a major English-language newspaper in Nepal and applies machine learning algorithms such as Logistic Regression and Naïve Bayes for sentiment classification. Text preprocessing techniques, including Porter’s Algorithm, are used to improve classification accuracy. Testing with training and test datasets demonstrated classification accuracies of 73% and 80% for Logistic Regression and Naïve Bayes respectively. The system provides users with article links and emoticons representing the detected sentiment, promoting a more mindful news-reading experience.

This project focuses on image enhancement using Convolutional Neural Networks (CNNs) to improve the quality and clarity of noisy images. Image enhancement techniques are applied to reduce noise, sharpen details, and improve brightness, resulting in cleaner and more understandable outputs. The system utilizes CNN architecture along with the naïve gradient descent learning algorithm to process digitized images and generate enhanced visual results. By leveraging deep learning techniques, the project demonstrates the effectiveness of CNNs in handling image restoration tasks and producing robust outputs even in noisy conditions. The system highlights the growing importance of machine learning and image processing in improving digital image quality for practical applications.

This project focuses on developing a Nepali Speech Recognition System capable of converting spoken Nepali language into text. Speech recognition is a complex task due to variations in accents, dialects, contextual meanings, and environmental noise. While significant advancements have been made in major languages, automated speech recognition for Nepali remains in its early stages. To address this gap, the project utilizes Natural Language Processing (NLP) tools such as CMUSphinx, PocketSphinx, SphinxTrain, and SphinxBase to recognize and process Nepali speech. The system was trained using a custom dataset containing 33 Nepali sentences spoken by two individuals and achieved approximately 85% accuracy. This application has strong potential to assist disabled individuals and non-literate Nepali speakers by enabling speech-to-text document creation, demonstrating the practical value of speech recognition technology in the Nepali language context.

This project focuses on automatic image colorization, a challenging image processing task that converts grayscale images into realistic colored images without requiring user intervention. Two fully automated approaches were implemented and evaluated. The first approach utilizes Support Vector Regression (SVR) combined with Markov Random Fields (MRF) to predict color channels for grayscale pixels. The second approach employs a Convolutional Neural Network (CNN) based deep learning model to predict color values in the LAB color space before converting them into RGB images. Experimental analysis using approximately 200 trained datasets showed that the CNN-based approach generated more realistic and accurate colorizations, especially for images containing multiple objects. The outputs were evaluated using histogram comparison techniques such as Correlation, Chi-square, Intersection, and Bhattacharyya distance. This project demonstrates the effectiveness of deep learning techniques in advanced image restoration and enhancement tasks.

This project presents a Nepali Sign Language Recognition system designed to bridge the communication gap between the Deaf community and hearing individuals. The system uses a 2D Convolutional Neural Network (CNN) to automatically recognize hand gestures representing Nepali alphabets and numbers. Image preprocessing techniques such as grayscale conversion, thresholding, edge detection, and contour detection are applied to extract meaningful hand gesture shapes before feeding them into the neural network. The CNN architecture includes convolutional, pooling, nonlinear, and fully connected layers, with ReLU activation used to introduce non-linearity. The model is trained on a dataset of 56,400 images covering 37 alphabets and 10 numerals, with data augmentation techniques such as horizontal flipping and inclusion of blank images to improve robustness. These enhancements significantly improved model performance, increasing accuracy from 82.45% to 92.46%. The system demonstrates the effectiveness of deep learning in sign language recognition and promotes accessible communication technology.