Enabling and Enforcing Social Distancing For COVID-19 Using Fully Convolutional YOLO Neural Networks
Author: Hasib Md Abid Bin Farid, Uchchhash Sarkar, Susmita Karmaker, Tanmoy Sarkar, Md. Jahedul Islam.
Abstruct: The novel coronavirus, first identified in Wuhan, China in December 2019, rapidly spread across the globe and was declared a pandemic by the World Health Organization (WHO) in March 2020. By June 2021, the virus had affected approximately 1.7 billion people worldwide and resulted in nearly 3.8 million deaths. Achieving widespread immunity is essential to ending the pandemic, and vaccination remains the safest and most effective means of accomplishing this. While numerous pharmaceutical companies have developed vaccines—many already approved by various governments—the challenge now lies in ensuring equitable distribution so that populations in all countries, not only wealthier ones, receive adequate protection. Until full vaccination coverage is achieved, preventive measures such as mask wearing, cleaning and disinfection, and social distancing remain crucial. Among these, social distancing is one of the most effective strategies for reducing transmission. This paper introduces a deep learning–based crowd monitoring framework designed to support social distancing. The system employs the YOLO object detection model to identify individuals in video sequences using bounding-box localization. After detecting people in a given scene, the system transmits the processed data to a web-based platform, allowing users to view crowd density information in real time. With this tool, individuals can choose less crowded routes to reach their destinations, thereby supporting adherence to social distancing guidelines. The proposed framework demonstrates the potential of artificial intelligence to assist public health efforts during pandemics.
Recognition of Bangla Handwritten Characters using Convolutional Neural Network
Author: Hasib Md Abid Bin Farid, Partha Sanjoy Dev, Nur Mohammad Antor, Mohammad Yusuf Harun, Md. Shariar Tanvir Sead.
Abstract: Recognition of handwritten characters of Bangla language, the vowels and numerical digits to be exact, is the prior focus of the work upon which this book is created. Keeping in mind the vast potential applications and practical usability as well as scalability to a huge extent, we initially generated our working interest in the field of character recognition which led our way to getting a satisfactory testing accuracy in detection of vowels and numerical digits of Bangla language. Following the workflow of other contemporary researches in the similar field of studies and after getting some particularly thorough researches on the relatively more popular approaches that others used previously, the coworkers of this project unanimously agreed upon using the Deep Learning methods, in which, the algorithms and model of Convolutional Neural Networks i.e. CNN to be exact, to train and test the partially unique, handwritten dataset. A thoroughly narrative theoretical description of the models and activation layers of the used methodology, the complete process of creating a fully unique dataset and its preprocessing steps of going to mere handwritten characters to usable csv file input and the acquired accuracy results of testing as well as its comparison to other probable results using different models were shown in detail in this paper.
Breast Cancer US Image Segmentation and Classification using Mask Region Based Convolutional Neural Network
Author: Hasib Md Abid Bin Farid, Zubair Hasan Chowdhury, Bashir Mahmud Talukder, Sohana Nazneen, Md. Sadman Islam Dipro.
Abstract:Â Breast cancer is one of the most significant global health-related problems among women with the level of occurrence that is steadily increasing. The pathogenesis of this malignancy is supported by a great number of risk factors, including chronological age, family history, genetic predisposition, hormonal processes, the influence of ionizing radiation, obesity, alcohol use, and comprehensive reproductive history. Therefore, timely and accurate diagnosis is required to assure effective treatment procedures and improvement of prognostic outcome. This paper presents a deep learning-based system to segment and classify breast cancer in ultrasound images. In particular, the Mask Region-Based Convolutional Neural Network (Mask R-CNN) is applied to outline and target tumorous areas that can be found in the sonographic images. The main objective of the work is to determine the subtype and anatomical location of breast cancer in the provided ultrasound images. It is expected that the results of this study will support the future research work and also enhance the diagnostic measures in breast cancer detection to improve the potential results.
Segmentation of Skin cancer using U-Net Neural Network
Author:Â Hasib Md Abid Bin Farid, Istiaque Ahmed, Supria Roy, Atika Nawar, Mostofa Habib.
Abstract:Â The unchecked growth of aberrant cells in the skin leads to skin cancer that can spread to other parts of the body hence causing a great risk to the survival of the patient. These abnormal cells are then identified in time, and thus the most important thing. The recent technological advancements, especially the use of machine learning algorithms on ultrasound imaging have improved the efforts of early detection. However, natural noise and irregularity of the ultrasound images make it difficult to outline the affected regions and this requires the creation of powerful segmentation models and methods. In the current study, we have used U-Net, a dedicated deep-learning network designed to be used in biomedical image segmentation, to extract domains with skin-cancer in ultrasound images. On the basis of platforms like Google Colaboratory and Spyder, we performed segmentation on various test images and compared the results with the respective ground-truth masks in the application of various segmentation methods. We have shown that U-Net achieves good segmentation results. In addition, we have found a positive correlation between accuracy of segmentation and image size as well as training epochs and also batch size, which means that the greater the above parameters the higher the accuracy of segmentation. This paper highlights the effectiveness of U-Net to ultrasound based segmentation of skin-cancer and provides useful information regarding parameter optimization in future studies.