Embedded Files
Research
Research
Pneumonia—a serious lung infection caused by bacteria or viruses—continues to be a global health concern, often diagnosed using chest X-rays. Our published study explores how deep learning can revolutionize this diagnostic process, offering fast, reliable, and automated detection of pneumonia from chest radiographs. Using a curated dataset of 5,863 chest X-ray images, including 4,273 pneumonia cases and 1,583 normal cases, we trained and validated a hybrid deep learning model capable of accurately distinguishing between healthy and affected lungs. The images were meticulously categorized into training, testing, and validation sets to ensure optimal performance and generalization.
Our approach involved preprocessing enhancements such as morphological operations and histogram equalization to amplify diagnostic features within the X-ray scans. These steps, combined with the power of deep convolutional neural networks, significantly improved detection accuracy and reduced false negatives. This research highlights how AI-driven diagnostic tools can assist radiologists and healthcare professionals in early, accurate identification of pneumonia—ultimately supporting faster treatment and better patient outcomes.
Colon cancer is a growing global health concern, especially in South Asia, where incidence rates are rising rapidly. Early detection is critical to improving survival rates, making the development of accurate diagnostic tools essential. This study aims to leverage advanced AI and machine learning techniques to detect colon cancer effectively. Using a dataset of 6,000 raw colonoscopy images categorized into four classes—Normal, Ulcerative Colitis, Polyps, and Esophagitis—we applied comprehensive image preprocessing to enhance data quality. We then implemented three state-of-the-art deep learning models along with a customized architecture based on InceptionResNetV2 integrated with specialized layers. Our custom model achieved a superior accuracy of 95.88%, outperforming other well-known models such as VGG19, InceptionV3, and ResNet50, which also showed promising results. Furthermore, we conducted a comparative analysis against recent related studies to validate our model’s performance, demonstrating its potential as a reliable tool for early and accurate colon cancer diagnosis.
Early detection enables accurate diagnosis of skin diseases, which is vital for effective treatment and management. Traditional diagnostic methods depend heavily on dermatological expertise, often being time-consuming and prone to human error. This study proposes an automated skin disease detection system leveraging deep learning to improve diagnostic accuracy and accessibility. A diverse dataset containing images of skin conditions such as chickenpox, measles, monkeypox, and healthy skin was used. Image preprocessing techniques—including resizing, stretch contrast, gamma correction, and augmentation—were applied to enhance image quality and ensure dataset uniformity. Several deep learning architectures were developed and evaluated, including DenseNet-121, ResNet-50, EfficientNet-B0, and a novel hybrid model that combines the strengths of these networks to achieve superior classification performance.
- Bangla Handwritten Character Recognition using Machine Learning(Thesis)
Recognizing handwritten Bangla characters is a complex challenge due to the language’s intricate script, diverse writing styles, and numerous classes of characters. In this thesis, we developed a machine learning-based system that can accurately identify Bangla handwritten characters, opening the door for powerful OCR applications, educational tools, and digitization of handwritten documents in Bengali.
We used a curated dataset of handwritten Bangla characters representing vowels, consonants, and compound characters. After extensive preprocessing—such as noise removal, binarization, and image normalization, we trained various machine learning models, including Support Vector Machines (SVM), k-NN, and Convolutional Neural Networks (CNNs). Among these, CNNs showed the best performance, capturing the fine-grained visual patterns unique to each Bangla character. The final model achieved high accuracy and robustness against variations in handwriting styles. This work contributes to language technology development in Bangla, making it easier to build intelligent applications that can read, process, and understand handwritten content in one of the most spoken languages in the world.
Projects
Projects
A Step Toward Assistive Tech for the Visually Impaired
This innovative project focuses on building a Smart Blind Stick using the Arduino Uno, designed to improve mobility and safety for visually impaired individuals. The device uses ultrasonic sensors to detect nearby obstacles and provides real-time audio feedback through varying alarm intensities—ranging from strong to medium—based on the distance of the obstacle.
Key components include:
Arduino Uno as the core controller
Ultrasonic sensors for obstacle detection
Buzzer to deliver adaptive alarm signals
Battery-powered, portable design
When an object is detected at a close range, the system triggers a strong buzzer alert, while medium-range objects generate a milder warning, enabling users to navigate their environment more safely. This project demonstrates a practical application of embedded systems and IoT to solve real-world problems and aims to empower visually impaired individuals through affordable, accessible technology.
A full-stack online shopping platform built with HTML, CSS, JavaScript, PHP, and MySQL. Features include user login, product browsing, shopping cart, secure checkout, and an admin dashboard for managing products and orders. Designed to provide a smooth, responsive shopping experience.
An automated door control system that uses sensors and microcontrollers to enable hands-free access. Designed for security and convenience, it detects presence and opens or closes the door automatically, enhancing safety and ease of use.
This project involves designing an intelligent street lighting system that automatically switches lights ON at night and OFF during the day using light sensors (LDR) connected to a microcontroller. It helps conserve energy by ensuring the lights operate only when needed, reducing electricity costs and minimizing human intervention. The system can also be expanded with motion sensors to increase brightness when pedestrians or vehicles are nearby, further optimizing power usage and enhancing safety.
Fare Fair is an intuitive Android app designed to help passengers quickly find the exact fare for their travel destinations across various routes. By entering the start and end points, users receive accurate fare information, eliminating confusion and ensuring transparency. The app supports multiple transport modes and regularly updates fare data to provide reliable, real-time pricing for commuters.
This project utilizes an Arduino microcontroller paired with smoke and temperature sensors to detect the presence of fire or smoke in an environment. When a potential fire is detected, the system triggers an audible alarm and visual indicators to alert nearby individuals immediately. This automated alert system enhances safety by providing early warnings and reducing response time in emergency situations.
This project automates the monitoring of over 480 CCTV cameras using PowerShell scripting. The system runs scheduled checks on camera status and connectivity, generating quick reports within minutes. By automating these routine inspections, it ensures timely detection of any camera downtime or faults, significantly improving surveillance reliability and reducing manual effort.
Page updated
Google Sites
Report abuse