Browsing by Author "Mostakim, Moin"

Now showing 1 - 20 of 111

A comparative study of car image generation quality using DCGAN and VSGAN
(BRAC University, 2022-05) Shayer, Mirza Ahmad; Anjum, Nafisha; Mim, Sushana Islam; Chowdhury, Md. Abu Sajid; Preoshi, Noshin Nanjiba Islam; Mostakim, Moin
In today’s modern society, image generation (synthesis) has a great number of uses in various tasks. Image generation is used in crime forensics, improving image quality and generating better images. In 2014, a scientific breakthrough occurred in the machine learning community when Ian Goodfellow and his colleagues introduced the GAN (Generative Adversarial Network). Ever since then, GANs have become a more popular concept in the scientific community. Even today, GANs are being used, utilized and upgraded. This thesis is a comparative study of two GANs used for generating images of cars- DC-GAN (Deep Convolution) and VS-GAN (Vehicle Synthesis). The study will determine which of the two is better suited to generate high quality images of cars. We will train both GANs using the same dataset. The dataset consists of about 16185 Google images of random cars, 8144 for training and another 8041 for testing. The dataset is already preprocessed and split. We will compare the GANs training times, losses, accuracies and pictures generated, showing how well they perform. We will run all the GANs for 40 epochs in both training and testing. We will compare the CGAN, DCGAN, VSGAN, WGAN and WGAN-GP, to see which performs the best. We have used K-Nearest Neighbors, Regression and Random Forest Classifier to calculate the accuracies of all the GANs. We have displayed the results in tabular and graphical formats. We believe this will improve GAN research by providing an excellent comparison between the GANs and determine which is better suited for the given task. We also hope to improve the models further in the future and make an even more in depth comparison between the GAN architectures.
A domain and noise adversarial bird tune classification pipeline using deep neural network
(BRAC University, 9/29/2022) Riya, Aparna Sarker; Roy, Arpita; Fahim, Md. Abrar; Tasnim, Zarin; Islam, Rakibul; Mostakim, Moin; Reza, Md Tanzim
Birds are an important category of animals that ecologists keep track of utilizing autonomous recording units as a key indication of environmental health. Because of the consequences of climate change and the rising number of endangered species, many experts suggested developing an animal species recognition system to help them in specialized research. Researchers can improve their ability to assess the state of biodiversity and its patterns in crucial ecosystems by precise sound detection and categorization, which is supported by machine learning, allowing them to better support global conservation efforts. However, producing analysis outputs with high precision and recall remains a difficulty. Due to a lack of appropriate methods for efficient and accurate extraction of interest signals, the vast bulk of data remains unexplored (e.g., bird calls). Moreover, due to strong source-domain specific features and artificial/natural noises, these acquired raw data create different distributions in datasets. So, to ensure a generalized feature learning, domain adaptation [1] techniques will be implemented in this work to make the networks familiar towards both acquisition sensor noises and background noises without having to do intensive dataset specific augmentations. We used 3 popular and powerful DNN models, including CNN, VGG19 and ResNet50. Out of them, for the bird species classification task VGG19 achieved the best accuracy of 96.02% in testing and 94.01% in training. To the best of our knowledge, this will guide towards convenient and deployable in real life models which will allow future works into the pipeline to ensure better coverage.
A machine learning approach to credit default prediction and Individual credit scoring
(BRAC University, 2018) Rahman, Md. Jaber; Ahmed, Hasib; Alam, A. N. M. Sajedul; Majumdar, Mahbub Alam; Mostakim, Moin
In our country the credit scoring system is not in practice yet so as for our undergrad thesis, we have taken upon the challenge of delivering a model well equipped with machine learning techniques to predict loan defaults. Here our main goal is to forecast credit defaults using machine-learning techniques and so we developed a model to output a target score, known as “credit score” which will describe the trustworthiness of an individual for getting a loan. We trained and tested this model based on ‘German credit data’, which was modified later on. We have Figured out 37 features based on which the data were taken and then after feature selection, we narrowed the number to 23 only by means of feature selection. Then again after thorough observations we analyzed the dataset with different models like Logistic Regression, FLDA, Naïve Bayes, Decision tree, Gradient Boosting tree, Random Forest etc. After that we made a scoring format using weights derived from information gains and also depending on their correlations, which will ensure the assigning of credit score to an individual. Later on we predicted who should receive loan on basis of the scores generated and this prediction was done using a decision tree.
A non-classical approach to recommender system for competitive programmers
(BRAC University, 2017-07) Chowdhury, Tahanima; Rafiq Ullah, Ahmed; Maqsud Ul Anwar, MD.; Mostakim, Moin
To practice solving programming problems from online automated judge is very essential for competitive programmers since this helps develop critical thinking required for solving programming challenges in various programming contests. Competitive programmers use a lot of their time searching for programming challenges in automated online judges whose level of difficulty are within their reach to solve. This is very time consuming and at times the programmers miss out some of the available programming problems. This issue has not been addressed in any previous works so we have come up with a non-classical approach to recommending challenges to competitive programmers.
Ahar24: a modern restaurant food delivery management system
(BRAC University, 2024-11) Iqbal, Syed Bayes; Mostakim, Moin
The rapid digital transformation has significantly impacted the food and restaurant industry, leading to the emergence of online food delivery systems. This research presents the development of Ahar24: A Modern Restaurant Food Delivery Management System, a comprehensive platform designed to streamline food delivery operations and enhance user experience. The system integrates advanced features such as personalized food recommendations powered by AI, a smart menu search with natural language understanding, dynamic pricing and discounts, and real-time order prediction and optimization. These features cater to both customers and administrators, ensuring seamless interactions, efficient order management, and data-driven decision-making. For customers, the platform offers a user-friendly interface for browsing menus, placing orders, and accessing tailored food suggestions based on preferences and past behavior. Administrators benefit from robust management tools, including insights into customer behavior, trending items, and demand forecasting. The system is implemented using cutting-edge technologies such as React.js, Node.js, and MongoDB, ensuring scalability and efficiency. While still under development and testing, Ahar24 demonstrates the potential to redefine modern restaurant operations and food delivery services by incorporating innovation, user-centric design, and advanced analytics.
AI thesis helper
(BRAC University, 2024-12) Shayer, Mirza Ahmad; Mostakim, Moin
In the modern world, Artificial Intelligence (AI) has become popular. Machine Learning (ML), is another popular term used on the internet, it is when AI is applied to a system which allows it to learn. Deep Learning (DL) is another popular term that became very common in recent times, it is the application of ML that utilizes algorithms and trains models. In 2020, Generative Pre-trained Transformer (GPT), came to the mainstream and had become highly successful. In just another year it became a game changing tool in the scientific landscape. Using GPT, various chatbots are designed. Chat-GPT in particular is by far the most used and most useful model of GPT. The model is useful in generating replies to various questions with high accuracy and usefulness. Can be used to help users with almost practically any question. In a few years this model will develop to help people even further. As such, this project - AI Thesis Helper, is one such implementation of further helping students to aid their thesis writing. Many students struggle to write their thesis, as such this chat-bot can help students with various thesis tasks like idea generation. This project aims to assist students in their writing. Project challenges and objectives are elaborated. Prior works are looked into along with reasons for choosing GPT over other models. Requirements, pros, cons and feasibility - both economic and technical along with model diagrams are dealt with. Details on the dataset are discussed. Fine-tuning of the model is elaborated with all details like imports, hyperparameters and training loop etc. Front-end, integration and backend are discussed along with UI/UX components. Implementation shows the project workings and how well it performs. To evaluate the performance, the following metrics are utilized - model loss during training, precision, recall, F1 score, perplexity and spelling, grammar with overall writing quality. Along with a comparison of generated text, made with two other AI chatbots - ChatGPT and Perplexity AI for the same text input. Future updates are also described.
An analysis of audio classification techniques using deep learning architectures
(BRAC University, 2020-10) Imran, Mohammed Safwat; Rahman, Afi a Fahmida; Tanvir, Sifat; Kadir, Hamim Hassan; Iqbal, Junaid; Mostakim, Moin
Failure to classify audio data with high efficiency causes major setbacks in audio processing, voice recognition and noise cancellation. In order to find the best possible neural network models for audio classi cation, this paper shows the steps in the experiments done on our newly designed CF Model and CFClean Model in both CNN and RNN, and compares the results with some existing models such as DCNN and Piczak-CNN. To get a clear view on the consistency of the results, three di erent datasets have been experimented on: UrbanSound8k, FSDKaggle2018 and ESC-50. This paper also sheds light on which dataset performs best in terms of train and test accuracy and loss percentage. Moreover, this paper also dives deep into the reasons behind particular models and datasets performing better than the others. Finally, this paper shows what influence envelope function, normalization, segmentation, regularization techniques and dropout layers have in the overall progress.
An analysis on the effects of parenting style on offspring’s behavior using machine learning
(BRAC University, 12/4/2022) Akter, Nasrin; Mostakim, Moin; Reza, MD Tanzim
Parents are usually the most important person for a human being as they encourage and support an offspring’s physical, emotional, social, and intellectual development from infancy to maturity. An individual faces various challenges as they grow up. Proper parenting plays a prominent role in handling and abating those challenges. This paper aims to show various consequences on the attachment style and handling of depression, anxiety, stress, anger due to different types of parenting style. These consequences of parenting styles are to be figured out in an automated way so that one can acknowledge these factors on their own and bring various positive changes to their parenting. The term ”parenting style” refers to a collection of tactics that have various effects on children. These methods can have an impact on children’ minds that lasts long into adulthood, both positively and negatively. This research makes use of machine learning algorithms in order to differentiate between various parenting styles through various aspects of their life such as stress, anxiety, depression, attachment style, anger management etc. The lack of publicly accessible data prompted us to compile my own data set, which consisted of 2206 survey responses from students(school, college, university). Afterward, the survey data was stored and pre-processed. Then, machine learning algorithms such as Decision Tree, XG-BOOST, KNN, Support Vector Machine and Random Forest are utilized to detect parenting style by analyzing the effects of parenting on their offspring and the accuracy of these models are 84.70%, 76.71%, 87.30%, 87.30% and 85.185% sequentially.
An analytic framework on food trend of Dhaka city using social media reviews and real time data set
(BRAC University, 12/5/2016) Alif, Redwanul Alam; Mostakim, Moin
I worked with sentiment analysis and supervised machine learning to find pattern and predict the current food trend and business The current food trend of Dhaka city depends on the type of food, the location of the restaurants, their overall services, and the taste of the available food items. In my analysis, I have shown the public sentiment on food trend depending on two (2) major types of food: Local or Deshi and Foreign food items. I have also analyzed the food trend depending on the reviews and rating people shared on social media based on their location according to the basis of taste, price, and service. Through the graph that I achieved from the collected data, I can also analyze the food business of Dhaka city depending on the food types and location of the restaurants. In my research I have used the data of reviews and ratings from different social media groups. Based on these reviews sorted according to locations like Dhanmondi, Banani, Gulshan, Khilgaon and Shyamoli I have analyzed the taste rating, price rating and service ratings. I was able to successfully show the food trend and restaurant business pattern of these locations based on the public reviews using necessary graphs.. This is how by applying supervised machine learning I have analyzed the public sentiment on restaurant and food trend of Dhaka city.
An enhanced CNN model for classifying skin cancer
(BRAC University, 2022-01) Haider, Kazi MD Minhajul; Dhar, Mondira; Akter, Fahima; Islam, Sadia; Shariar, Syed Ragib; Hossain, Muhammad Iqbal; Mostakim, Moin
Unrepaired deoxyribonucleic acid in skin cells causes skin cancer by generating genetic abnormalities or mutations, rising day by day. Detecting and diagnosing skin cancer in its early stages is expensive and challenging, giving superior treatment options. Given the severity of these issues, researchers have generated a set of early classification techniques for skin cancer. Skin cancer is diagnosed and segregated from melanoma by looking at the symmetry, color, size, shape, and other features of lesions. While there are various computerized approaches for classifying skin lesions, convolutional neural networks (CNNs) have been demonstrated to exceed standard practices. Moreover, CNNs are a type of deep learning that has been prominent in various fields, including medical imaging. Multiple machine learning libraries have been used in this paper. Also, we have used five pre-trained models such as Inception V3, VGG-19, VGG-16, Efficient Net B7, ResNet 50 models and presented our proposed model for skin cancer classification using the HAM10000 dataset, which is an enormous skin cancer dataset. Following that, each competent model’s image detection categorization accuracy is evaluated by comparing and assessing. This research reports a maximum accuracy of 85.25% for Inception V3 models within five pre-trained models and maximum accuracy of 90.55% for our proposed model. In terms of image detection, our experimental configuration shows that our proposed model can attain the best classification accuracy rather than the other five pretrained models. Our findings are helpful in providing a comprehensive comparison and analysis of many neural networks in the categorization of skins cancer.
An improved image registration technique for shape reconstruction
(BRAC University, 2017) Roy, Biprojit; Oishe, Geetanjali; Ali, Md. Haider; Mostakim, Moin
Image registration is a very useful technique in the field of Image processing. It has many potential applications for shape and angle estimation. Image registration is a basically aligning two or more images into one common co-ordinate system in order to monitor subtle changes between them. The image registration algorithms rotate ortranslate the source images and determine the relative transformation with the corresponding target image. In this research we are going to reconstruct the shape of a scene/object from different images. The viewpoint of the images will be arbitrary and independent of any available data set. From set of source images we are going to combine the images using image registration techniques to find the estimated shape. We believe this research will be useful regarding the studies of shape reconstruction.
An Iris detection and recognition system to measure the performance of e-security
(BRAC University, 4/21/2016) Anwar, A.M. Shahed; Biswas, Rubel; Mostakim, Moin
Biometric is a system to identify the individual human by extracting distinguishable features from that particular person. Among many other biometric systems the iris recognition system is most accurate one right now since it has a high recognition rate. This thesis is proposing a system with four major division in the process: Segmentation, Normalization, Feature encoding and Matching. At early stage, Histogram equalization is used on the input image and to detect the objects present in the image, Canny Edge Detection model is employed. Inner circular boundary and center of the pupil in the Iris region is detected by using Hough Transformation. A circle is drawn with the help of Mid-Point Circle drawing algorithm which center is as same as pupil center and hence, outer circular boundary of the iris region can be detected. For the normalization, Daugman’s Rubber Sheet model is used and in the feature encoding process, instead of Gabor filter to extract feature from the iris image, Log-Gabor filter is used in this thesis since it has non-zero DC component advantage over Gabor filter. Last but not the least, Hamming Distance is used to compare two binary iris template for matching purpose.
An optimized predictor for patient health records while ensuring HIPAA compliance
(BRAC University, 2026-01) Das, Stanley Matthew; Alam, Ashiqul; Alvi, Arif Jawad; Mostakim, Moin; Nasim, Hamim Ibne; Tanvir, Sifat
The increasing digitization of healthcare has led to predictive analytics becoming an essential tool for early risk detection and personalized patient care. This project introduces an optimized predictor for Patient Healthcare Records. This is a microservices-driven, AI-based system architected to analyze patient data while maintaining HIPAA (Health Insurance Portability and Accountability Act), ensuring scalability through Dockerized deployment. The system functions through three main phases: (1) Data processing through Optical Character Recognition (OCR), which extracts text and refines patient data from medical records; (2) Health Risk Prediction utilizing a Hidden Markov Model (HMM) for sequential health analysis and Neural Networks for predictive modeling; and finally, (3) Secure storage and Recommendations where the predictions are organized in a structured PostgreSQL database and accessed via a web/mobile platform built with HTML and CSS. This design guarantees effective, privacy-conscious, and AI-enabled healthcare analytics, delivering real-time insights for healthcare professionals and providing them with a streamlined, scalable, and secure method for health risk prediction, supporting proactive medical decision-making.
Application of deep convolutional neural network in breast cancer prediction using Digital Mammograms
(BRAC University, 2022-01) Al Mamun, Rafsan; Rafin, Gazi Abu; Alam, Adnan; Sefat, MD. Al Imran; Bin Ashraf, Faisal; Mostakim, Moin
Cancer, a diagnosis so dreaded and scary, that its fear alone can strike even the strongest of souls. The disease is often thought of as untreatable and unbearably painful, with usually, no cure available. Among all the cancers, breast cancer is the second most deadliest , especially among women. What decides the patients’ fate is the early diagnosis of the cancer, facilitating subsequent clinical management. Mammography plays a vital role in the screening of breast cancers as it can detect any breast masses or calcifications early. However, the extremely dense breast tissues pose difficulty in the detection of cancer mass, thus, encouraging the use of machine learning (ML) techniques and artificial neural networks (ANN) to assist radiologists in faster cancer diagnosis. This paper explores the MIAS database, containing 332 digital mammograms from women, which were augmented and preprocessed, and fed into a custom and different pre-trained convolutional neural network (CNN) models, with the aim of differentiating healthy tissues from cancerous ones with high accuracy. Although the pre-trained CNN models produced splendid results, the custom CNN model came out on top, achieving test accuracy, AUC, precision, recall and F1 scores of 0.9362, 0.9407, 0.9200, 0.8025 and 0.8572 respectively while having minimal to no overfitting. The paper, along with proposing a new custom CNN model for better breast cancer classification using raw mammograms, focuses on the significance of computer-aided detection (CAD) models overall in the early diagnosis of breast cancer. While a diagnosis of breast cancer may still leave patients dreaded, we believe our research can be a symbol of hope for all.
Aspect based opinion mining on restaurant reviews
(BRAC University, 2023-09) Tonmoy, Sazid Hasan; Ahmed, Faiyaj Bin; Sarkar, Madhurjya; Bashar, Mehejabin Binta; Ahmed, Rafi; Sadeque, Farig Yousuf; Mostakim, Moin
The way businesses are operating have changed due to the explosion of the internet. Social media has an increasing number of reviews as people are keen to express their opinions based on their experiences. Online reviews have become a precious asset in various disciplines such as intelligent marketing and decision-making.The number of reviews for a well-liked product might reach thousands. This makes it challenging for a prospective buyer to go through them and make up their minds. In order to overcome this challenge, a machine-learning system is needed. Aspect based Opinion mining can be used to extract the aspects from the reviews, then we can analyze the nature of the reviews and recommend them to all the customers. We plan to classify reviews about a target entity as positive, negative and neutral so that readers of the reviews do not have to go through all the reviews but instead can focus on functional items and applicable suggestions. This thesis is specifically focused on reviews in the domain of restaurants. This study extends our knowledge of online reviews by taking into account users’ wants and anticipating their future behavior. Several distinct evaluative linguistic nuances shed light on internet reviews. Using an assortment of models on generated benchmark datasets, we will also empirically show the efficacy of our strategy and show that the new techniques (or modified versions) are superior to, or at least on par with, state-of-the-art methods.
Aspect-based sentiment analysis using SemEval and Amazon datasets
(BRAC University, 2017) Hasib, Tamanna; Rahin, Saima Ahmed; Mostakim, Moin
Sentiment analysis has become one of the most important tools in natural language processing, since it opens many possibilities to understand people’s opinions on different topics. Aspect-based sentiment analysis aims to take this a step further and find out, what exactly someone is talking about, and if he likes or dislikes it. Real world examples of perfect areas for this topic are the millions of available customer reviews in online shops. There have been multiple approaches to tackle this problem, using machine learning, deep learning and neural networks. However, currently the number of labelled reviews for training classifiers is very small. Therefore, we undertook multiple steps to research ways of improving ABSA performance on small datasets, by comparing recurrent and feed-forward neural networks and incorporating additional input data that was generated using different readily available NLP tools.
Assistive Guideline of Categorization for Competitive Programming Problems
(BRAC University, 2021-06) Dhrubo, Najmus Sakib; Islam, Md Samiul; Rafi, Tahmid Ul Islam; Mostakim, Moin
Programming is a very useful skill nowadays. Programming contests give people the opportunity to increase their programming skills. By solving programming contest problems contestants can increase not only their programming skills but also their mathematical and algorithmic knowledge. The competitive programming problems are presented in problem statements. Sometimes they are presented in the form of a story or sometimes directly. To solve the problem contestants must read the problem statement carefully. The problems can be of many categories. We have tried to classify number theory and graph theory problems. At first, we collected data from competitive programming problem statements. Then we used different machine learning algorithms such as fully connected neural network, naive bayes classifier, support vector machine on the data to predict if the category of the problem is either number theory or graph theory. With such machine learning approaches we achieved test accuracy of about 72%, 75% and 74%.
Automated fabric color prediction
(BRAC University, 2023-05) Muhtashima, Fawzia; Maksurah, Fawzia; Jahan, Sifat E; Mostakim, Moin
"This paper focuses on addressing some challenges faced by colorists and explores various approaches to predict fabric color changes after dyeing processes. It empha- sizes the importance of color prediction in the textile industry and proposes suitable models that can effectively carry out color prediction tasks based on given recipes. By implementing such predictive models, the textile industry can improve efficiency, reduce labor-intensive practices, and enhance the overall quality control process. The methods used in this study are supervised machine learning techniques, in- cluding multiple linear regression, decision tree, random forest, and neural network. Among these models, the most appropriate one is selected and further optimized using feature engineering techniques to improve accuracy"
Automated short answer scoring
(BRAC University, 12/14/2016) Mithu, M. Rayhan Ahmed; Robin, Syed Mohammad Moinul Islam; Kamal, Moumita; Mostakim, Moin; Abdullah, Matin Saad
In our thesis we have worked to analyses text short answers then predict the score accordingly by using different extracted features. In our research we have used around 1700 data for each dataset and are scored by two different humans provided by the Hewlett foundation available in Kaggle. We have used different NLP techniques to process the data in order to use it to the classifiers. Sckit was used to implement the algorithms of the different classifiers. The data were divided into two different data sets, one of them was the training set and the test set. The training set data was used to train the different classifiers afterwards the test set data was given to the classifiers to predict the score. The predicted score was compared with the score given by the humans to find the efficiency and accuracy of the different classifiers.
Automatic Bengali image captioning using efficientNet-transformer network and vision transformer
(BRAC University, 2023-01) Kabir, Muhammad Khubayeeb; Labonno, Anindita; Amin, Sofia; Tahsin, Fariha; Rahman, Md. Khalilur; Mostakim, Moin
The task of image captioning is a complex process that involves generating textual descriptions for images. This technology is extremely beneficial for a wide range of applications, such as assisting people with visual impairments, monitoring surveil lance systems, content generation, image indexing, and automatic annotation of images for producing data for training AI-based image generation models. Much of the research done in this particular domain, especially using transformer models, has been focused on English language. However, there has been relatively little research dedicated to the context of the Bengali language. This study addresses the lack of research in the context of Bengali language and proposes a novel approach to auto matic image captioning that involves a multi-modal, transformer-based, end-to-end model with an encoder-decoder architecture. Our approach utilizes pre-trained Ef ficientNet Transformer Network. To evaluate the effectiveness of our approach, we compare our model with a Vision Transformer that utilizes a non-convolutional en coder pre-trained on ImageNet.The two models were tested on the BanglaLekhaIm ageCaptions dataset and evaluated using BLEU metrics.