Enhancing Poultry Farm Productivity Using IoT-Based Smart Farming Automation System

Short Description: Our study presents a smart farming system utilizing Internet of Things technology, specifically designed for poultry agriculture. This system concentrates on meticulous monitoring of environmental conditions and analytics to present clear guidance and suggestions for maximizing production. Key components consist of instant surveillance, mechanized distribution of feed and water, and a specialized smartphone app. With the aid of analytic-driven responses and efficient automatization, this system proposes a practical option to conventional agricultural methods.

Publisher: IEEE

Date of Publication: February 27, 2024

DOI: 10.1109/ICCIT60459.2023.10441525

Performance Analysis of Tesseract and EasyOCR for Bangla Optical Character Recognition on the Novel Bangla-CrossHair Dataset

Short Description: In this study, we put together a novel dataset on Bangla OCR for the purpose of fine tuning and analyzing different OCR Engines. There has been a limited collection of dataset in bangla language. Our dataset is not limited to documents, but also extends to image instances such as blurred, clear, torn and tilted images.The benchmarking was performed using metrics like: Character Error Rate, Word Error Rate, Levenshtein Distance, Precision, Recall, and F1 Score. Our study concludes that while Tesseract OCR outperforms EasyOCR on reading text from documents, it fails to capture text from challenging instances effectively without image preprocessing which is where EasyOCR excelled. Also, Tesseract OCR had faster inference times in all cases which is expected due to the architectural differences between the models. [ACCEPTED]

Conference: 2025 3rd International Conference on Intelligent Systems, Advanced Computing and Communication (ISACC)

Towards Developing an Automated Breast Cancer Diagnosis Tool from Mammograms Using Deep Learning

Description: In this work, we have implemented the image segmentation architecture to segment the ROI regions of the mammograms along with their sizes to fit the processed images. We have developed a web app using Flutter as frontend and Flask as backend with Firebase as user authentication. We dedicate this tool to the people of the underdeveloped countries such as Bangladesh where advanced medical diagnostic tools are scarce. We had submitted our tool in the Microsoft Imagine Cup 2024 and progressed up to the quarter finals of the competition.

Research Area: Computer Vision, Deep Learning, Medical Image Processing, AI Diagnostic Tool, Health Informatic

Supervisor:
Dr. Md. Mostafa Kamal Sarker BSc. M.Eng Ph.D.
Visiting Fellow, University of Oxford, UK
Lead AI Research Scientist, Technovative Solutions Limited, UK

DART-UNet: L2-Norm Position Encoding and Residual Cross-Cov Attention in a Dual-Attentive Transformer UNet for Fetal Head Imaging Segmentation

Description: : In my research, I developed DART-UNet, a novel deep learning framework that integrates a ResNet101 encoder, a novel Dual Attention Blocks, and Vision Transformers to improve fetal head ultrasound segmentation. To enhance feature representations, I have developed a Spatial Position Attention Block (SPAB) for capturing spatial dependencies and a Residual Cross-Covariance Attention Block (RCCAB) for modeling inter-channel relationships. Our model uses L2-norm position encoding and residual cross-covariance attention. We have trained our model on the Large-Scale Annotation Dataset for Fetal Head Biometry, it uses a composite Dice-Focal loss for optimized performance.

Research Area: Computer Vision, Deep Learning, Image Processing, Image Segmentation

Status: Submitted in IEEE Access

Supervisor:
Shadman Sakib
Graduate Teaching Assistant
University of Maryland Baltimore County, Maryland, US

Privacy-Preserving Dynamic Session Management for Federated Learning in a Web-Based Healthcare AI Framework

Description: : Our research is of a framework that has a privacy-preserving dynamic session management system for a web-based federated learning platform in healthcare. It enables secure, decentralized AI model training while ensuring compliance with data privacy regulations and supporting resource-limited hospitals with scalable, user-friendly tools.

Research Area: Federated Learning, Cryptography, Deep Learning

Real-time Respiratory Rate Signal Waveform Estimation from Video Stream

Description: : Developing a novel methodology for estimating respiratory rate (RR) signal waveforms from video streams. The video is preprocessed using Eulerian magnification; a region of interest (ROI) is then selected using a pose estimator, which is Mediapipe and optical flow with RAFT; finally, the video is modified for input to a diffusion model for training. The trained model will be able to estimate RR signal waveform from a video stream.

Research Area: Computer Vision, Deep Learning, Data Augmentation, Signal Processing