Ahmed Faizul Haque Dhrubo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Ahmed Faizul Haque Dhrubo has a total of 25 co-authors. Here are the top co-authors by number of shared publications:
Mohammad Abdul Qayum
Department of Electrical and Computer Engineering, North South University, Dhaka, 1229, Bangladesh
3 shared publicationsMohammad Ashrafuzzaman Khan
Department of Electrical and Computer Engineering, North South University, Dhaka, 1229, Bangladesh
2 shared publications- Acramul Haque Kabir
North South University
1 shared publication - Al Imran
Department of Electrical and Computer Engineering, North South University, Dhaka, Bangladesh
1 shared publication Amit Chakraborty
North South University
1 shared publication- Aritra Islam Saswato
Department of Electrical and Computer Engineering, North South University, Dhaka 1229, Bangladesh
1 shared publication
Latest Publications
- Open Access
RadMultiBench: A Reproducible Dual-Task Benchmark Exposing the Limits of Pre-Trained Vision-Language Models Across Thoracic and Breast Imaging
We present RadMultiBench, a multimodal benchmark of 6,148 medical images paired with expert radiological impressions across five categories—Breast Cancer, Lung Benign, Lung Malignant, Pneumonia, and Tuberculosis—targeting two core clinical tasks: automated report generation and disease classification. To enable reliable, repeatable experimentation, we provide a reproducible pipeline for data ingestion, de-identification, split management, and standardized reporting of text and classification metrics, facilitating dataset versioning and continuous benchmark testing. We compare three architectures: (1) a classical CNN-RNN with Bahdanau attention (ResNet-LSTM), (2) a general-domain vision-language (V-L) prefix-projection model (CLIP-GPT), and (3) a domain-specific V-L prefix-projection model (BioViL-T). Results reveal a sharp performance divide: the ResNet-LSTM attains near-perfect classification (99.5% accuracy) and is the only model to generate coherent reports (BLEU-4 = 0.021), whereas both prefix-projection models catastrophically fail generation (BLEU-4 ≤ 0.0001); moreover, the domain-specific BioViL-T also fails classification (11.5% accuracy). From a data-engineering angle, the pipeline is productionizable with lineage, feature registries, and continuous integration to guard against silent regressions when updating models or data; from a data-science angle, the benchmark provides baselines and analyses that foreground feature alignment and architectural suitability as first-order concerns, inviting ablations on multimodal objectives, retrieval-augmented encoders, and uncertainty-aware decoding. Overall, RadMultiBench highlights that state-of-the-art status in general V-L domains does not translate automatically to high-stakes clinical tasks, and it furnishes the tooling and evidence base needed to develop robust, clinically viable multimodal systems.
Open AccessRadMultiBench: A Reproducible Dual-Task Benchmark Exposing the Limits of Pre-Trained Vision-Language Models Across Thoracic and Breast Imaging
We present RadMultiBench, a multimodal benchmark of 6,148 medical images paired with expert radiological impressions across five categories—Breast Cancer, Lung Benign, Lung Malignant, Pneumonia, and Tuberculosis—targeting two core clinical tasks: automated report generation and disease classification. To enable reliable, repeatable experimentation, we provide a reproducible pipeline for data ingestion, de-identification, split management, and standardized reporting of text and classification metrics, facilitating dataset versioning and continuous benchmark testing. We compare three architectures: (1) a classical CNN-RNN with Bahdanau attention (ResNet-LSTM), (2) a general-domain vision-language (V-L) prefix-projection model (CLIP-GPT), and (3) a domain-specific V-L prefix-projection model (BioViL-T). Results reveal a sharp performance divide: the ResNet-LSTM attains near-perfect classification (99.5% accuracy) and is the only model to generate coherent reports (BLEU-4 = 0.021), whereas both prefix-projection models catastrophically fail generation (BLEU-4 ≤ 0.0001); moreover, the domain-specific BioViL-T also fails classification (11.5% accuracy). From a data-engineering angle, the pipeline is productionizable with lineage, feature registries, and continuous integration to guard against silent regressions when updating models or data; from a data-science angle, the benchmark provides baselines and analyses that foreground feature alignment and architectural suitability as first-order concerns, inviting ablations on multimodal objectives, retrieval-augmented encoders, and uncertainty-aware decoding. Overall, RadMultiBench highlights that state-of-the-art status in general V-L domains does not translate automatically to high-stakes clinical tasks, and it furnishes the tooling and evidence base needed to develop robust, clinically viable multimodal systems.
- Open Access
- 8 September 2026
- Published by Elsevier in SSRN Electronic Journal
Automated Vision-Based Corrosion Inspection for Steel Infrastructure Using Lightweight Adaptive Deep Learning
In this work, we investigate task-adaptive asymmetric loss calibration for lightweight corrosion segmentation using a CBAM-enhanced MobileNetV2 U-Net. First, we compare DiceCE with a reference Focal Tversky-based loss and observe that the latter produces a strongly precision-oriented behavior, substantially reducing corrosion recall and overall overlap. Motivated by this limitation, we propose the Industrial Defect Asymmetric Loss (IDAL), which systematically adjusts the asymmetric weighting of the Focal Tversky component to identify a task-specific precision–recall operating point. The selected configuration is further evaluated using multi-seed experiments and five-fold cross-validation. IDAL achieves comparable corrosion IoU and Dice to DiceCE (0.498 vs. 0.497 and 0.665 vs. 0.664), while improving precision (0.637 vs. 0.583), Boundary IoU (0.114 vs. 0.104), and HD95 (45.924 vs. 48.879). This benefit is accompanied by reduced recall (0.695 vs. 0.774). A limitation is that the evaluation is conducted on a single corrosion dataset.
- 8 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAutomated Vision-Based Corrosion Inspection for Steel Infrastructure Using Lightweight Adaptive Deep Learning
In this work, we investigate task-adaptive asymmetric loss calibration for lightweight corrosion segmentation using a CBAM-enhanced MobileNetV2 U-Net. First, we compare DiceCE with a reference Focal Tversky-based loss and observe that the latter produces a strongly precision-oriented behavior, substantially reducing corrosion recall and overall overlap. Motivated by this limitation, we propose the Industrial Defect Asymmetric Loss (IDAL), which systematically adjusts the asymmetric weighting of the Focal Tversky component to identify a task-specific precision–recall operating point. The selected configuration is further evaluated using multi-seed experiments and five-fold cross-validation. IDAL achieves comparable corrosion IoU and Dice to DiceCE (0.498 vs. 0.497 and 0.665 vs. 0.664), while improving precision (0.637 vs. 0.583), Boundary IoU (0.114 vs. 0.104), and HD95 (45.924 vs. 48.879). This benefit is accompanied by reduced recall (0.695 vs. 0.774). A limitation is that the evaluation is conducted on a single corrosion dataset.
- Open Access
- 19 May 2026
- Published by Springer Nature in Research Square
Reg2Bangla: An End-to-End Regional Speech Standardization
This paper presents a complete approach for transcribing twenty regional Bangladeshi dialects into standard Bangla text. We fine tuned the tugstugi bengaliai regional ASR Whisper medium model, which is a fine tuned variant of the Whisper model from OpenAI trained on external datasets. After this, we further fine tuned the model on a corpus of three thousand three hundred fifty dialectal audio recordings using the annotated label texts in the training set. We also applied post processing with an n gram KenLM language model and used KV cache optimization to achieve faster inference for the audio input. The proposed methodology tackles major challenges that arise from regional variation, pronunciation differences, and vocabulary shifts across linguistic communities in Bangladesh. The proposed pipeline shows strong performance on the evaluation set and demonstrates that a combination of pretrained multilingual speech models and targeted fine tuning, supported by post-processing techniques, can effectively manage the complexity of Bangladeshi dialectal speech.The corresponding code has been made publicly available at https://github.com/sazzadadib/shobdotori-bangla-dialect-ai-televerse.
- 19 May 2026
- Published by Springer Nature in Research Square
Open AccessReg2Bangla: An End-to-End Regional Speech Standardization
This paper presents a complete approach for transcribing twenty regional Bangladeshi dialects into standard Bangla text. We fine tuned the tugstugi bengaliai regional ASR Whisper medium model, which is a fine tuned variant of the Whisper model from OpenAI trained on external datasets. After this, we further fine tuned the model on a corpus of three thousand three hundred fifty dialectal audio recordings using the annotated label texts in the training set. We also applied post processing with an n gram KenLM language model and used KV cache optimization to achieve faster inference for the audio input. The proposed methodology tackles major challenges that arise from regional variation, pronunciation differences, and vocabulary shifts across linguistic communities in Bangladesh. The proposed pipeline shows strong performance on the evaluation set and demonstrates that a combination of pretrained multilingual speech models and targeted fine tuning, supported by post-processing techniques, can effectively manage the complexity of Bangladeshi dialectal speech.The corresponding code has been made publicly available at https://github.com/sazzadadib/shobdotori-bangla-dialect-ai-televerse.
- Open Access
- 24 April 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (4) , e0346977
PlantaNet and PlantaNetLite: Efficient and explainable multi-crop plant disease classification via transformer benchmarking and custom lightweight CNNs
Plant disease diagnosis based on visual symptoms is crucial for preventing yield loss; however, deployment in practical settings remains challenging due to inter-class similarity, background noise, and limited computational resources. This study presents a plant disease classification framework evaluated on a curated multi-crop dataset aggregated from multiple publicly available repositories, comprising 51 disease and healthy classes. The dataset includes approximately 45,000 original images that were expanded through controlled augmentation during training to improve generalization. We benchmark eight ImageNet-pretrained tiny vision transformer architectures trained for up to 50 epochs. Among these, CAFormer-s18 achieved strong validation performance but with increased computational overhead. To enable efficient and computationally lightweight solutions, we design two fully customized convolutional neural networks: PlantaNetLite (1.28M parameters) and PlantaNet (2.58M parameters). After hyperparameter optimization and full 100-epoch training, PlantaNet achieved 99.37% validation accuracy and 99.66% test accuracy with a compact model size (9.85 MB) and moderate computational cost, while PlantaNetLite achieved a best validation accuracy of 99.22% under further parameter reduction. Qualitative Grad-CAM and Grad-CAM++ analyses provide insight into the regions influencing model predictions. Overall, the proposed models demonstrate competitive accuracy while maintaining computational efficiency, highlighting their potential suitability for resource-constrained deployment scenarios.
- 24 April 2026
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 21 (4) , e0346977
Open AccessPlantaNet and PlantaNetLite: Efficient and explainable multi-crop plant disease classification via transformer benchmarking and custom lightweight CNNs
Plant disease diagnosis based on visual symptoms is crucial for preventing yield loss; however, deployment in practical settings remains challenging due to inter-class similarity, background noise, and limited computational resources. This study presents a plant disease classification framework evaluated on a curated multi-crop dataset aggregated from multiple publicly available repositories, comprising 51 disease and healthy classes. The dataset includes approximately 45,000 original images that were expanded through controlled augmentation during training to improve generalization. We benchmark eight ImageNet-pretrained tiny vision transformer architectures trained for up to 50 epochs. Among these, CAFormer-s18 achieved strong validation performance but with increased computational overhead. To enable efficient and computationally lightweight solutions, we design two fully customized convolutional neural networks: PlantaNetLite (1.28M parameters) and PlantaNet (2.58M parameters). After hyperparameter optimization and full 100-epoch training, PlantaNet achieved 99.37% validation accuracy and 99.66% test accuracy with a compact model size (9.85 MB) and moderate computational cost, while PlantaNetLite achieved a best validation accuracy of 99.22% under further parameter reduction. Qualitative Grad-CAM and Grad-CAM++ analyses provide insight into the regions influencing model predictions. Overall, the proposed models demonstrate competitive accuracy while maintaining computational efficiency, highlighting their potential suitability for resource-constrained deployment scenarios.
- Open Access
Dynamic Sensitivity Filter Pruning using Multi-Agent Reinforcement Learning For DCNN's
Deep Convolutional Neural Networks have achieved state of the art performance across various computer vision tasks, however their practical deployment is limited by computational and memory overhead. This paper introduces Differential Sensitivity Fusion Pruning, a novel single shot filter pruning framework that focuses on evaluating the stability and redundancy of filter importance scores across multiple criteria. Differential Sensitivity Fusion Pruning computes a differential sensitivity score for each filter by fusing the discrepancies among gradient based sensitivity, first order Taylor expansion, and KL divergence of activation distributions. An exponential scaling mechanism is applied to emphasize filters with inconsistent importance across metrics, identifying candidates that are structurally unstable or less critical to the model performance. Unlike iterative or reinforcement learning based pruning strategies, Differential Sensitivity Fusion Pruning is efficient and deterministic, requiring only a single forward-backward pass for scoring and pruning. Extensive experiments across varying pruning rates between 50 to 70 percent demonstrate that Differential Sensitivity Fusion Pruning significantly reduces model complexity, achieving over 80 percent Floating point Operations Per Seconds reduction while maintaining high accuracy. For instance, at 70 percent pruning, our approach retains up to 98.23 percent of baseline accuracy, surpassing traditional heuristics in both compression and generalization. The proposed method presents an effective solution for scalable and adaptive Deep Convolutional Neural Networks compression, paving the way for efficient deployment on edge and mobile platforms.
Open AccessDynamic Sensitivity Filter Pruning using Multi-Agent Reinforcement Learning For DCNN's
Deep Convolutional Neural Networks have achieved state of the art performance across various computer vision tasks, however their practical deployment is limited by computational and memory overhead. This paper introduces Differential Sensitivity Fusion Pruning, a novel single shot filter pruning framework that focuses on evaluating the stability and redundancy of filter importance scores across multiple criteria. Differential Sensitivity Fusion Pruning computes a differential sensitivity score for each filter by fusing the discrepancies among gradient based sensitivity, first order Taylor expansion, and KL divergence of activation distributions. An exponential scaling mechanism is applied to emphasize filters with inconsistent importance across metrics, identifying candidates that are structurally unstable or less critical to the model performance. Unlike iterative or reinforcement learning based pruning strategies, Differential Sensitivity Fusion Pruning is efficient and deterministic, requiring only a single forward-backward pass for scoring and pruning. Extensive experiments across varying pruning rates between 50 to 70 percent demonstrate that Differential Sensitivity Fusion Pruning significantly reduces model complexity, achieving over 80 percent Floating point Operations Per Seconds reduction while maintaining high accuracy. For instance, at 70 percent pruning, our approach retains up to 98.23 percent of baseline accuracy, surpassing traditional heuristics in both compression and generalization. The proposed method presents an effective solution for scalable and adaptive Deep Convolutional Neural Networks compression, paving the way for efficient deployment on edge and mobile platforms.