Baharul Islam builds thinking systems.

Baharul Islam, Nasim Ahmad, Mehbub Alam, Sk Mahmudul Hassan, Sudip Chakraborty, Khaled A. Mahmoud, S. Aljlil, Ruhul Amin Hazarika

Abstract

Underwater infrastructures such as pipelines, ship hulls, and offshore platforms are critical to marine operations but are highly vulnerable to biofouling, structural corrosion, and vegetation overgrowth, leading to increased maintenance costs and environmental hazards. However, visual inspection of these structures remains challenging due to low visibility, uneven lighting, and complex textured surfaces that limit the effectiveness of both traditional and purely deep learning-based approaches. In this work, we introduce AquaFusionNet, a hybrid defect classification framework that seamlessly integrates embeddings from four state-of-the-art pre-trained backbones (EfficientNet-B0, ResNet-50, SENet-50, and Vision Transformer) with complementary traditional descriptors including colour histograms, histogram of oriented gradients (HOG), local binary patterns (LBP), edge density, and gradient statistics via a trainable attention module. This attention mechanism dynamically weights each feature channel, allowing the model to emphasise the most informative cues while preserving fine-scale details under variable turbidity and illumination. We evaluate AquaFusionNet on a curated dataset of 2,228 underwater images spanning three defect categories, with 445 images held out for testing. Our model achieves 98.43% accuracy, 98.18% precision, 97.79% recall, a 96.06% intersection-over-union, and a 97.97% F₁-score, outperforming eleven strong baselines, including ResNet-152 and EfficientNet-B0, by a substantial margin. These results demonstrate AquaFusionNet's robustness and generalisability, paving the way for real-time, automated underwater inspection systems that can significantly enhance operational safety and reduce maintenance costs across marine industries.

Baharul Islam, Nasim Ahmad, Ferdous Ahmed Barbhuiya, Kuntal Dey

Abstract

This paper presents a system for automated subject tagging in a bilingual academic setting. Our approach leverages a novel burst attention mechanism to enhance the alignment between article and subject embeddings, derived from a large cross-lingual subject corpus. By employing a margin-based loss with negative sampling, our resource-efficient model achieves competitive performance in both quantitative and qualitative evaluations. Experimental results demonstrate average recall rates of 32.24% on the full test set, along with robust performance on specialized subsets, making our system well-suited for large-scale subject recommendation tasks.

Nasim Ahmad, Baharul Islam, Mehbub Alam, Sudipta Majumder

Abstract

Road accidents attributed to driver drowsiness impose a severe social and economic burden worldwide. We present a hybrid vision–based framework that combines compact embeddings from a Vision Transformer (ViT, specifically NV-DINOv2) with a lightweight multi-layer perceptron (MLP) classifier to detect driver fatigue from facial cues and eye dynamics. By decoupling high-capacity visual feature extraction (ViT) from a small task-specific head (MLP), the proposed system delivers state-of-the-art accuracy while remaining fast enough for real-time deployment on commodity hardware. The model attains 99.25% test accuracy together with 99% precision, 99% recall, and a 99% F1 score substantially outperforming MobileNetV2, VGG16, and VGG19 baselines. These results demonstrate the effectiveness of ViT embeddings for robust, low-cost driver-monitoring applications.

Abstract

Abstract