Face Recognition Using ArcFace Embeddings and Traditional Machine Learning Techniques

This paper presents a facial recognition pipeline combining ArcFace-based facial embeddings with traditional machine learning classifiers to identify individuals with associated metadata. The pipeline includes DNN-based face extraction and Haar Cascade frontal-face alignment. ArcFace embeddings are used as feature inputs, achieving 100% accuracy on Logistic Regression, KNN, SVM, and Ensemble models on ORL and Yale datasets, and 93.1% accuracy on Logistic Regression over a mixed dataset of 12,372 images across 210 labels — delivering a deployable, resource-efficient recognition system.

 Index Terms: Face Recognition, ArcFace, Machine Learning, Logistic Regression, SVM, Identity Classification, Profiling System, Model Comparison

Bangla Fake News Detection Using BanglaBERT and Ensemble Machine Learning Models

This paper addresses the proliferation of fake news in the linguistically under-resourced Bangla language. A comprehensive, scalable detection pipeline is presented using BanglaBERT — a Bangla-optimised transformer — benchmarked against SVM, Logistic Regression, Naive Bayes, XGBoost, LightGBM, and LSTM models. Advanced preprocessing (stopword removal, lemmatization, balanced sampling) is applied across multiple datasets, yielding strong results. The proposed ensemble model achieves 93.09% accuracy, contributing a reproducible framework for NLP research in low-resource languages.

 Index Terms: Fake News Detection, BanglaBERT, NLP, Ensemble Learning, LSTM, Low-Resource Languages, Bangla NLP

Multi-Modal Explainable Deep Learning for Air Quality Index Classification

An attention-based multi-modal deep learning system that fuses environmental image features (EfficientNet-B3) with numerical sensor data (TabNet) to classify AQI into six categories. The fusion layer dynamically learns to weight image vs. sensor modality per prediction. Built-in Grad-CAM and attention distribution plots provide explainability. Enables AQI estimation from smartphone photos and local sensor readings — making air quality monitoring more accessible in sensor-sparse regions.

Hybrid-Solar-Forecaster: Multi-Horizon Renewable Energy Prediction

A high-precision multi-horizon solar energy forecasting framework combining Temporal Fusion Transformer (TFT) and N-BEATS branches, fused via an attention-based weighting layer. The pipeline applies Kalman filtering and wavelet transforms for denoising, and Isolation Forests for anomaly removal. Optimised with quantile and Huber loss functions, the model achieves lower RMSE and MAE compared to standalone architectures — supporting grid stability and energy trading efficiency for the Alice Springs solar facility.

Fintech Fraud Detection with Hybrid Deep Learning and Behavioral Analytics

An end-to-end credit card fraud detection system using a PyTorch hybrid neural network with embedding layers for categorical variables and a dense numerical branch. Geospatial Haversine distance analysis, rolling behavioral features (transaction velocity), and target statistical encoding are incorporated. The model maximises F1-score via dynamic thresholding, addressing the severe class imbalance inherent in financial transaction datasets — enabling real-time integration into payment gateways.

Political Sentiment Analysis Using Transformer Models and Bangla NLP

A robust NLP pipeline for sentiment classification of political and social text, built on a BERT-family transformer (BanglaBERT variant) fine-tuned for specialised discourse datasets. Integrates the CSE BUET normalizer for language-specific preprocessing of Bengali scripts, SentencePiece BPE tokenization, and GPU-optimised training (NVIDIA T4). Applicable for political sentiment tracking, brand monitoring, and automated customer service routing.

OcuScan: Multi-Modal Ocular Disease Recognition with Graph Attention Networks

A hybrid multi-label classification system for detecting eight ocular disease categories (Normal, Diabetes, Glaucoma, Cataract, AMD, Hypertension, Myopia, and Other) from fundus images combined with patient metadata. A Swin Transformer encodes visual features while a Graph Attention Network (GAT) captures disease co-occurrence dependencies. Grad-CAM++ heatmaps provide explainability for clinical trust, enabling early screening in areas lacking ophthalmic specialists.

Deep Learning for Air Pollution Classification from Environmental Images

An image-based deep learning study focused on classifying levels of air pollution directly from environmental photographs. The research explores CNN-based and transfer-learning architectures trained with atmospheric augmentation strategies — including RandomFog and GaussNoise — to simulate varying pollution and visibility conditions, aiming to provide a low-cost, camera-based complement to traditional sensor networks.

Birds Classification Using Transfer Learning and Fine-Tuned CNN Architectures

A fine-grained bird species classification study leveraging transfer learning on pre-trained convolutional neural network backbones. The research investigates feature extraction and full fine-tuning strategies, data augmentation pipelines, and cross-architecture comparisons to achieve high-accuracy species identification — contributing to biodiversity monitoring and ecological research through accessible, automated visual recognition.