all projects

Machine Learning

Multi-Modal Emotion Recognition

This project focuses on constructing a multi-modal emotion recognition system with facial and audio features. It employs subspace-based feature fusion methods, including z-score normalization and Canonical Correlation Analysis (CCA), to combine facial expression and audio features. SVM classifiers are trained using spatiotemporal features from 50 videos of 5 participants and evaluated using the remaining data. LOSO cross-validation is implemented for reliable performance estimation.

ScipyPCASVMCCANumPyLOSO
Machine Learning

Speech to Emotion Recognition

In this project, speech from ten speakers is utilized to construct an emotion recognition system distinguishing between happy and sad states. Basic prosodic features and Mel-Frequency Cepstral Coefficients (MFCC) are extracted from the speech recordings. SVM classifiers are trained using a subset of the data and evaluated using the remaining samples. The goal is to compare the effectiveness of simple prosodic features with MFCC features in recognizing emotions.

signalscipySVM
Computer Vision

Facial Expression Analysis

This project involves facial expression analysis and feature classification tasks. The region of interest (i.e., facial image) is extracted using face tracking, face registration and face crop functions. Basic spatiotemporal features (i.e., LBP-TOP features) are extracted using LBP-TOP. For classificaiton of the extracted features, SVMs are trained. 50 videos from 5 participants are used to train the emotion recognition using spatiotemporal features. The rest of the data (50 videos) is used to evaluate the performance of the trained recognition system.

Scikit-LearnMatplotlibScipydlib
Natural Language Processing

Information Retrieval from Texts

This project involves tasks such as analyzing corpora datasets, generating frequency plots, calculating word lengths, examining modal words and implementing information retrieval systems. Additionally, it explores web scraping from Wikipedia, constructing tf-idf matrices and evaluating search results using tf-idf representations.

NLTKBeautifulSoapScikit-Learn
Natural Language Processing

Text to Emotion Detection

The project focuses on exploring and analyzing emotion detection from texts using various techniques including natural language processing, manual categorization, word embeddings, and similarity matching. It involves preprocessing text, identifying frequent keywords, reducing categories, comparing different categorization approaches and evaluating using metrics like token overlap and Pearson correlation.

GloveFastTextBERTSVM
Application

YouTube Statistics

This application extracts data from YouTube Data API v3 that scraps videos of a channel with tags and statistics. This application also has a scheduler that tracks changes in video statistics (viewCount, likeCount, favoriteCount and commentCount) and tags every 3 mintues and stores and updates them into the postgreSQL database.

DockerDjangoReactPostgreSQL
Application

Blood Donation Application

The blood donation application integrates three vital services: a PostgreSQL Database for secure data storage, a Django REST Framework for robust backend functionality including pagination and centralized logging followed by API testing and a React frontend with MaterialUI for a seamless user experience. This app facilitates efficient donor management, ensuring smooth operations and accessibility for both donors and recipients.

DockerReactDjangoPostgreSQL
Machine Learning

Network Intrusion Classification through Oultier Detection

We provided an approach for categorizing cyber-attacks effectively through outlier detection. We applied machine learning approaches including LR, SVM, AdaBoost, NB and KNN. We evaluated the efficacy of our approach on three network intrusion datasets (KDD Cup 99, CIC-IDS2017, and UNSW-NB15).

Isolation ForestDBScanKMeans
Deep Learning

Ventilation Pressure Prediction using Transformer

We implemented a hybrid model (BiLSTM+Transformer) with multi-headed attention mechanism and layer normaliztion to predict the pressure within a mechanical lung on any given time step, based on how much the inspiratory solenoid valve is opened.

Bi-LSTMTransformerPyTorch
Natural Language Processing

Sentiment Analysis on Twitter Data using Transformer

We implemented a classic transformer combining the multi-headed attention mechanism with the encoder-decoder, feed forward Convo1D and layer normalisation on a text-based Twitter dataset with the help of GloVe pre-trained embeddings.

TransformerMultiHeaded AttentionGloVePyTorch
Deep Learning

Linear Regression using PyTorch on Diabetes Dataset

We designed and implemented a PyTorch model tailored for a linear regression problem. The training process utilized the gradient descent algorithm, orchestrated within PyTorch’s framework, to iteratively adjust and optimize the model's parameters for accurate predictions.

Linear RegressionPyTorch
Deep Learning

Deep Neural Network with Fashion-MNIST

We delved into the fundamentals of machine learning by constructing a basic neural network (NN) from scratch. This hands-on approach allowed us to grasp the intricacies of backpropagation. The project also explored stochastic gradient descent (SGD) for iterative optimization and we applied various regularization techniques and hyperparameter tuning strategies to enhance our neural network's performance.

Stochastic Gradient DescentHyperparameter TuningNumPy
Deep Learning

Transfer Learning with MiniImageNet and ViT

We worked with the miniImageNet focusing on the train.tar file and later split the data as train, val and test datasets. We pre-trained three models - ResNet18, VGG16 and Vision Transformer (ViT) with the best optimizers on the Training Set of MiniImageNet and then evaluated on the Validation Set and tested on the Testing Set. We saved the checkpoints for each of the models used to later use them on EuroSAT-RGB dataset. We took 5 classes from EuroSat-RGB dataset then randomply took samples as instructed in the project instruction. We used this dataset for transfer learning on the model checkpoints we saved for VGG16, ResNet18 and Vision Transformer (ViT).

VGG16ResNet18Vision TransformerMiniImageNetEuroSAT-RGB
Natural Language Processing

Data Driven Understanding of Emotions from Bangla Texts

Since there was no corpus accessible for Bangla texts, we had to build our own corpus first, tokenize the data, extract the features from the texts, analyze the sentiment and classify them into six types of emotions - happy, sad, anger, fear, disgust, and surprise. More than 20K sentences were collected associating specific emotional characteristics labelled with six classes and we also developed four classifier models to detect emotion using different machine learning techniques - SVM, Naive Bayes, Logistic Regression and Decision Tree.

TokenizationStop-word RemovalStemmingClassification
Machine Learning

Dyslexia Prediction from Data Driven Game Analysis

This method uses predictive ensemble machine learning to reliably identify dyslexia early by transforming feature space and finding the most relevant features in a dataset using an enhanced Genetic Algorithm with a modified fitness function as feedback. To test the accuracy, multiple machine learning methods with cross-validation have been employed. Random Forest has the best accuracy (99.5%). As a feature selection method, it compares to Recursive Feature Elimination.

PCAGenetic AlgorithmRandom Forest
Machine Learning

Cervical Cancer Risk Prediction

This research has shown a higher prediction accuracy of cervical cancer after missing value imputation. We have implemented Linear Discriminant Analysis for dimensionality reduction & Adaptive Synthetic Sampling approach (ADASYN) to properly balance the dataset and got a better outcome. Using Isolation Forest algorithm, we have effectively identified outliers in cervical cancer datasets and removed them from the dataset.

ADASYNLDAIsolation ForestDecision Tree
Machine Learning

Data Driven Diagnosis of Heart Disease

EDA is demonstrated by plotting the analysis of Univariate and Multivariate analysis. A correlation heatmap is illustrated and several Machine Learning algorithms have been applied for the original dataset showing around 86% accuracy for Logistic Regression, XGBoost and Support Vector Classifier. For the feature selection, we used feature importances and Recursive Feature Elimination having 25 optimal features. After feature selection, we achieved the highest accuracy 86.67% for Logistic Regression.

Multivariate AnalysisLogistic RegressionSVMXGBoost
Artificial Intelligence

Handwritten Digits Detection

This model takes MNIST dataset to fit the data into most widely used machine learning algorithms - KNN, SVM, Random Forest and deep learning algorithm Deep Convolutional Neural Network (Layer 3) and acquired 98.70% for DCNN as compared to 97.91% using SVM, 96.67% using KNN, 96.89% using Random Forest Classifier.

CNNRandom Forest
Machine Learning

Visualizing Linear Regression Predicting Housing Price

A real estate company has a dataset containing the prices of properties. It wishes to use the data to optimise the sale prices of the properties based on important factors such as area, bedrooms, parking, etc. The company wants to identify the variables affecting house prices, e.g. area, number of rooms, bathrooms, etc. The model takes a linear model that quantitatively relates house prices with variables such as number of rooms, area, number of bathrooms, etc. and finds the accuracy of the model, i.e. how well these variables can predict house prices.

PythonNumPyLinear Regression
Web Development

Profile Summary Card from GitHub API

The web API creates cards on each request, currently with no cache, so cards displayed on the client is always up to date. The system gets valid GitHub profiles data and alerts if entered username is invalid.

React HooksGitHub API
Application

Star Matching Game

I have created this repository following a Pluralsight course for learning react. This game is an implementation of star match game where you have to select the sum of numbers/number that the total should result in the number of stars shown. Follow the game until all of the numbers are used. If you finish the game in 'n' seconds, then you will win or you will lose.

React Hooks
Web Development

FlaskBlog

A simple blog using Python (Flask), SQLAlchemy, HTML, CSS, JavaScript, Bootstrap. It's a starter project in Flask. Admin can add, delete or create any blogs. Users can also add blog posts and will be displayed to other users based on date & time and popularity. Users can also give feedback, see the latest posts, announcements and event calendars.

FlaskSQLiteSQLAlchemy
Web Development

Instagram Cloning

It allows users to edit and upload photos, add a caption to each of their posts. Each post by a user appears on their followers' Instagram feeds and can also be viewed by the public. Users can like, comment on and bookmark others' posts, add and view unlimited stories. Users can track down the followers and followings, the total number of views and which users have viewed their content.

LaravelMySQLJavaScript

copyright©2024 Moinul Islam all rights reserved