Thesis (Bachelor of Science in Computer Science)

Browse

End date: 2019Subject: search.filters.subject.Machine learning

Search Results

Now showing 1 - 10 of 12
  • Thumbnail Image
    Item
    Detection of handwritten text using convolutional neural network
    (BRAC University, 2019-04) Jasim, Rabib Bin; Mahin, Rokeya Sultana; Uddin, Jia
    Machine replication of human functions, like reading, is an ancient dream. However, over the last five decades, machine reading has grown from a dream to reality. We have tried to make it more obvious through a hand writing recognition system. This research paper describes a text-line extraction based method. It offers a new solution to traditional handwriting recognition techniques using concepts of Deep learning and computer vision. An image can have hand writing, typed letters, different characters and other images. Our intention is to detect all the characters and display them. Some images can also have unnecessary lines or unclear letters. This system will clear the picture through pre-processing system and will be able to identify the letters or characters. It will help people to identify any unclear messages. It will also avoid unnecessary images and will focus on the text only. Sometimes we want to ignore unnecessary advertisement images from the newspapers. Our system will do a great work for this. It will clear all the images and unnecessary lines etc. and will only display the text what people want to read.
  • Thumbnail Image
    Item
    Comparative study of X-ray and CT scan images for the detection of COVID-19 using deep learning
    (BRAC University, 2015-08) Niloy, Ahashan Habib; Shiba, Shammi Akhter; Fahim, S.M. Farah Al; Faria, Faizun Nahar; Rahman, Md. Jamilur; Parvez, Mohammad Zavid
    Coronavirus 2019 (in short, COVID-19), originated in the Wuhan province of China in December 2019, has been declared a global pandemic by WHO in March 2020. Since its inception, it’s rapid spread among nations had initially collapsed the world economy and the increasing death-pool created a strong fear among people as the virus spread through human contact. Initially doctors struggled to diagnose the increasing number of patients as there was less availability of testing kits and failed to treat people efficiently which ultimately led to the collapse of the health sector of several countries. To help doctors primarily diagnose the virus, researchers around the world have come up with some radiology imaging techniques using the Convo lutional Neural Network (CNN). While some of them worked on x-ray images and some others on CT scan images, none worked on both the image types. Thus there’s no way to know which image works better for a particular model. This, therefore, insisted us to perform a comparison between x-ray and CT scan images. Thus we came up with a novel CNN model named CoroPy which works for both the image types and shows that in 2 classes (normal and covid), CT scan images show a better accuracy and it is 99.17% whereas it is 95.73% for x-ray images. However, in the case of 3 classes (normal, covid and viral pneumonia), x-ray images show a better accuracy and it is 92.45% whereas it is 68.81% for CT scan images.
  • Thumbnail Image
    Item
    Prediction of acute lymphoid leukemia using Privacy Preserving Neural Network
    (BRAC University, 2019-12) Khilji, Ishfaque Qamar; Saha, Kamonashish; Shonon, Jushan Amin; Israq, Ragib; Hossain, Muhammad Iqbal
    In today’s world, machine learning has become a big factor. It not only needs to be helpful, but also accurate and precise prediction is required. Machine learning is now becoming a widely used mechanism and applying it in certain sensitive fields like medical and financial data has only made things easier, but it also brought some difficulty in data privacy and data security which will protect the complete implementation of cloud based machine learning for these aspects due to the law and ethical needs. In this project, to give proper solution, we have come up with the idea using concepts of CryptoNets and Neural Networks, where we will be able to convert the learned neural network with the encrypted data to Cryptonets and the data will be totally encrypted and this will prevent the chances of unencrypted data being available to everyone. In this method, the owner will send the encrypted data to the cloud first and will hold a private key which can be used to decrypt the data later on. The cloud will have no idea about the data there since it will be in encrypted form and any attempts to get data from the cloud will only give the encrypted form. However, applying neural network to the cloud will enable us to store the data and make predictions in encrypted form and also give back the encrypted data to the user. In this way, the cloud will have no idea about the actual data and after the prediction is made, it will give back the predicted data in the encrypted form. We were able to achieve an encrypted prediction of about 78% close to the validation accuracy amount we achieved when training our Neural Network model.
  • Thumbnail Image
    Item
    Financial factors analysis for acquisition premium and anticipation using extreme gradient boosting and deep recurrent neural network
    (BRAC University, 2019-12) Rayhan, Mohammad; Sultana, Samiha; Majid, Annur; Alam, Md. Golam Rabiul
    This study shows importance hierarchy of financial factors of corporations’ Goodwill and tries to foresee with popular machine learning and deep learning models. Financial engineering is using mathematical model to study financial behavior. Financial engineers are hired by investment banks, commercial banks, hedge funds, insurance companies, corporate treasuries, and regulatory agencies. It is vital for each of them to asses a company’s sustainability before any sort of investment. However, predicting sustainability is not deterministic. Therefore, corporate sustainability has become a mainstream business goal for stakeholders. Whether Quantitative finance impacts goodwill or has implicit insight can be a machine learning problem. Deep learning and machine learning are rapidly changing the financial services industry. Business leaders can now transform vast amounts of financial data into insightful predictions with the help of data science, creating significant savings in the bottom line. This thesis is concerned with investigating financial factors of a company’s Goodwill and also fits popular machine learning and deep learning models and evaluate goodness of fit. To aid the research, a comparison between the proposed models-XGboost and Deep LSTM are conducted.
  • Thumbnail Image
    Item
    N.P.K. based crop suggesting model
    (BRAC University, 2019-12) Iftekher, Asif; Ghosh, Tomalika; Alam, Md. Ashraful
    Bangladesh is a country having an area of 1, 47,570 square kilometers in which roughly 70.63 percent is agricultural lands [1]. As an agricultural country we are mostly dependent on soil. There are 3 most important nutrients in any soil, it's known as the primary macro nutrients: Nitrogen (N), Phosphorus (P), and Potas- sium (K). Each of the primary nutrients is very essential in plant nutrition, serving a critical role in growth and reproduction of the plant. The purpose of this project is to make a N.P.K. Based Crop Suggesting Model by using machine learning which will determine the best crop to grow in a particular soil based on some major crite- ria. This model will play a vital role in our agricultural sectors to fulfill the needs of our country by reaching the highest level of efficiency and ensure the best use of our arable lands.
  • Thumbnail Image
    Item
    Detection of early stages of Parkinson's disease by analyzing fMRI data and machine learning approaches
    (BRAC University, 2019-12) Neehal, Ahmed Hasin; Azam, Md. Nura; Islam, Md. Sazzadul; Hossain, Md. Ishrak; Parvez, Mohammad Zavid
    Parkinson's Disease is a progressive nervous system brain disorder which affects motor neuron loss control and movement coordination. Parkinson's symptoms are shown gradually and get worse over time. Its signs and symptoms can be different for everyone. There may be minor early signs and they may go unnoticed. Therefore, early detection of Parkinson's disease might significantly improve life style by giving proper treatment. Moreover, doctors may suggest regulating certain regions of your brain and improve the symptoms. In recent years, the use of Functional Imaging in neurodegenerative diseases has increased, with applications in basic pathophysiology research, support in determination, or evaluation of new medications. In our research we used fMRI data of eight early PD patients. Resting-state fMRI images were collected for analyzing the data and feature extraction. Time series data were generated for each subject based on voxel intensity. In addition, STFT was used to measure the time frequency function. Furthermore, SVM classifier was used for the classification and prediction of the early stage of PD. Using our proposed method, we have achieved 100% sensitivity, specificity, and accuracy considering seven subjects, however, one subject was exceptional whereas we have achieved 99.76% accuracy, 100% specificity and 99.53% sensitivity. Finally, this process is a well-structured model for predicting the early stages of PD. It may help to the doctors for diagnosis of the disease at its early stages and the patients should receive better treatment.
  • Thumbnail Image
    Item
    Detection of skin cancer using Convolutional neural network
    (BRAC University, 2019-10) Ahsan, Abu Sa-adat Mohamed Moon-Im Al; Alif, Shadman Monsur; Kibria, Junaid Bin; Gomes, Prince Elvis; Uddin, Jia
    One of the most common and fatal cancer in the universe is skin cancer which arise from skin of epidermis, the topmost layer of the skin, it can happen anywhere in the body. We can find out the cancer by early detection. Skin cancer detection is a time consuming process and very critical. So in clinical applications, the machine learning analysis of skin cancer is failed to give correct images for a model. In our paper we followed three pre-processing steps which are: a) removing the shadows from the image which is illumination correction processing, b) to find the border of the skin lesion in the segmentation part, c) feature extraction by doing the ABCD framework. Our thesis makes an attempt to implement the method of Convolutional Neural Network. Using this classification, we find out the best result in inception v3 which was trained on skin lesions and we got the accuracy of 82.4%. So, our primary focus of this thesis is to differentiate between cancerous and non-cancerous image. Then our goal is to reduce importance of one of the painful process in cancer detection which is known as biopsy. Biopsy is removing tissue from a body and later it goes to many laboratory tests.
  • Thumbnail Image
    Item
    Performance analysis of machine learning classi ers for detecting PE malware
    (BRAC University, 2019-12) Azmee, ABM.Adnan; Choudhury, Pranto Protim; Alam, Md.Aosaful; Dutta, Orko; Hossain, Muhammad Iqbal
    In this modern era of technology, securing and protecting one's data has been a major concern and needs to be focused on. Malware is a program that is designed to cause harm and malware analysis is one of the paramount focused points under the sight of cyber forensic professionals and network administrations. The degree of the harm brought about by malignant programming varies to a great extent. If this happens at home to a random person then that may lead to some loss of irrel- evant or unimportant information but for a corporate network, it can lead to loss of valuable business data. The existing research does focus on some few machine learning algorithms to detect malware and very few of them worked with Portable Executables (PE) les. However, we worked on the PE les and also for real-time computation, a client-server model was developed by using Flask to detect malware or benign. In this paper, we mainly focused on top classi cation algorithms and compare their accuracy to nd out which one is giving the best result according to the dataset and also compare among these algorithms. Top machine learning clas- si cation algorithms were used alongside neural networks such as Arti cial Neural Network, XGBoost, Support Vector Machine, Extra Tree Classi er, etc. The exper- imental result shows that XGBoost achieved the highest accuracy of 98.62 percent when compared with other approaches. Thus, to provide a better solution for this kind of anomalies, we have been interested in researching malware detection and want to contribute to building strong and protective cybersecurity.
  • Thumbnail Image
    Item
    Building a credit scoring model to assign a reference score based on credit transaction and relevant profile data
    (BRAC University, 2019-09) Islam, Saqib Al; Aziz, Rifah Sama; Ahmed, Aritra; Abida, Fauzia; Majumdar, Mahbub Alam
    A credit score is a numerical expression based on a level analysis of a person's credit files, to represent the creditworthiness of an individual. The credit score plays a major role in banks, financial institutions loaning money to individuals for their personal or business needs. This score is given based on factors such as personal information, assets, financial behavior and financial history. This system is not digitized or implemented yet in Bangladesh. So our aim is to build a reliable and robust credit scoring model which would help institutions like such to have an accurate reference score to rely on when validating a client. We were able to obtain an optimized model with an accuracy of( 93%). The model is based on CART(Classification and Regression Trees) using Gradient Boosting method(GBM). We also proposed a new hybrid model consisting of a two step architecture. The first one based on distributed Random Forests, the individual decision tree outputs of which was fed into a Deep Neural Network(DNN), and trained on to achieve marginally better results than using only Random Forest approach. Since, credit scoring an individual is a sensitive issue, it is not ethical to provide a score without proper justification. We conducted interpret-ability analysis on our model and generated visual representations of the criterion affecting the output of our model and provide necessary information to analyze the client efectively. Our results were conclusive and imitated the process of evaluating an individual precisely. The work- ow we proposed could be implemented in production to provide a concrete base for evaluation and prediction of defaulters. Simultaneously provide a detailed overview of the results obtained. This could help financial institutions immensely and help them save millions lost by default loans.
  • Thumbnail Image
    Item
    Bangla sign language recognition using leap motion sensor
    (BRAC University, 2019-08) Tan, Tamkin Mahmud; Mondol, Anna Mary; Nawal, Noshin; Ahmed, Sabbir; Uddin, Jia
    Sign language is used by hearing and speech impaired people to transmit their messages to other people but it is difficult for a regular people to understand this gesture based language. Instantaneous responses on sign language can significantly enhance the understanding of sign language. In this paper, we propose a system that detects Bangla Sign Language using a digital motion sensor called Leap Motion Controller. It is a sensor or device which can detect 3D motion of hands, fingers and finger like objects without any contact. A Sign Language Recognition system has to be designed to recognize a hand gesture. In sign language system, gestures are defined as some specific patterns or movement of the hands to give an expression. There has to be a library which includes all the datasets to match with the user given gestures. We have to compare the sequences of data we get from Leap Motion and our datasets to get an optimal result which is basically the output. It will then show the output as text in the display. For our system, we choose to use $P Point-Cloud Recognizer algorithm to match the input data with our datasets. This recognition algorithm was designed for rapid prototyping of gesture-based UI and can deliver an average over 99% accuracy in user-dependent testing. Our proposed model is designed in a way so that the hearing and speech impaired people can communicate easily and efficiently with common people.