Majida Kazmi

Automated diabetic retinopathy (DR) screening has seen extensive research progress, yet its clinical adoption remains limited, largely due to insufficient attention to model generalizability across diverse populations and imaging conditions. Existing studies often overlook the role of dataset biases, arising from demographic, acquisition, and preprocessing variations that undermine robustness in real-world settings. To address this critical gap, we propose a comprehensive generalizability analysis framework that systematically categorizes dataset biases, evaluates their impact on performance, and introduces robust inter-dataset comparison metrics. The framework is metric-agnostic and extensible to multi-class severity grading, with stage-wise BAG formulations introduced to support more detailed clinical evaluation. Using one primary (EyePACS) and nine secondary datasets, we assessed the generalizability of two distinct architectures: a convolutional neural network (MobileNetV2) and a transformer-based model (CvT). Performance was evaluated through intra-group measures (accuracy, sensitivity, specificity, and AUC) and a newly proposed Bias-Adjusted Generalization (BAG) index designed to quantify resilience to bias-induced domain shifts. Results demonstrate that CvT consistently outperformed MobileNetV2, achieving an average accuracy of 87% and a BAG index of 0.94, compared to MobileNetV2’s 78% accuracy and BAG index of 0.91. Gradient-based explainability visualizations further confirm that both models attend to clinically relevant retinal regions, with architectural differences in activation patterns consistent with their respective generalizability profiles. These findings highlight the inherent bias resilience of transformer architectures, while highlighting their computational demands as a barrier to deployment in low-resource environments. Importantly, the study suggests that integrating transformer-inspired mechanisms into lightweight CNNs could yield clinically scalable models with both efficiency and robustness. Our framework offers a standardized approach for bias-aware evaluation, providing actionable insights for developing equitable, generalizable, and resource-adaptable AI solutions for DR screening.

Automated detection of surface defects in steel is critical for ensuring product quality and operational efficiency in modern manufacturing. This study presents a unified framework for real-time steel surface inspection using five lightweight YOLO (You Only Look Once) nano architectures: YOLOv5n,...

Sign language recognition is a critical area within computer vision and assistive technology, promoting inclusive communication for individuals with hearing impairments. While various international sign languages have been widely studied, Pakistan Sign Language (PSL) remains underexplored. This chapter, per the authors, presents a real-time PSL recognition system using the YOLOv5 object detection framework. A custom dataset of 3,963 images, augmented to 9,522 using rotation, noise addition, and exposure adjustment, was developed using hand signs from students at the Ida Rieu School for the Blind and Deaf. The model achieved an accuracy of 94.67%, outperforming previous PSL systems in static sign classification. Comparative analysis shows improved generalization in real world conditions, particularly with complex backgrounds and partial occlusions. Unlike earlier approaches limited to controlled environments, this system demonstrates practical applicability in educational contexts. Future work includes optimizing for low-power edge deployment and dynamic sign recognition.

This paper presents an edge-deployable, real-time adaptive traffic light control system designed for complex and undisciplined urban traffic environments. The system integrates computer vision and Internet of Things (IoT) technologies to enable smart traffic management. A robust dataset was developed to capture diverse traffic conditions, including occlusions and unstructured vehicle movement. A lightweight vehicle detection model, based on (You Only Look Once) YOLO and deployed on a Jetson Xavier NX-powered roadside edge node (RSEN), enables real-time traffic monitoring with a mean Average Precision (mAP) of 90% and 74 frames per second (FPS). To improve traffic flow representation, we introduce a traffic density estimation method that incorporates Passenger Car Equivalent (PCE) factors, considering the varying impact of different vehicle types. The estimated density is used by an adaptive traffic light control (ATLC) algorithm, deployed on a resource-constrained Portenta H7 board, which dynamically adjusts signal timings using real-time data, starvation prevention, and tie-breaking mechanisms. The proposed system reduces traffic congestion at intersections by up to 33% and waiting times by 23% compared to traditional fixed-time controls. The proposed framework is cost-effective, scalable, and fully operational on edge devices, offering a practical solution for real-world smart city deployments.

Deploying deep learning–based optical character recognition (OCR) systems for low-resource, complex-script languages like Urdu remains a major challenge due to high computational costs, lack of annotated datasets, and limited hardware support for real-time applications. Existing FPGA-based OCR implementations have primarily focused on simplified datasets such as MNIST digits, limiting their generalizability to scripts like Urdu that exhibit extensive intra-class variability, contextual shaping, and diacritics. This study presents a hardware-accelerated Urdu OCR framework using a custom-designed Convolutional Neural Network (CNN) optimized for deployment on the Xilinx Zynq UltraScale+ MPSoC (ZCU104). The proposed CNN is trained on a novel large-scale dataset of 336,000 labeled images spanning 48 Urdu characters across 230 font styles. Compared to MNIST-based FPGA implementations, our approach addresses significantly higher script complexity while achieving a classification accuracy of 96.73% (FP32) and 94.06% (INT8). Hardware-aware quantization and deployment using the Vitis AI toolchain enabled 75% model compression with minimal accuracy loss, achieving real-time inference of 0.189 ms per character and 4,886.95 FPS, while consuming only 1.32 W. Benchmarking against CPU and GPU platforms confirmed substantial improvements in speed and energy efficiency. This work establishes a high-performance, scalable, and energy-efficient FPGA-based OCR framework for Urdu and sets the foundation for extending such solutions to other cursive, low-resource languages like Arabic, Pashto, and Persian.

Energy management is important for both consumers and utility providers. Utility providers are concerned with identifying and reducing energy wastage and thefts. Consumers are interested in reducing their energy consumption and bills. In Pakistan, residential and industrial estates account for nearly 31,000 MW of the maximum total demand, while the transmission and distribution capacity has stalled at about 22,000 MW. This 9000 MW gap in demand and supply, as reported in 2022, has led to frequent load shedding. Although the country now has an excess generation capacity of about 45,000 MW, the aging transmission and distribution network cannot deliver the requisite power at all times. Hence, electricity-related problems are likely to continue for the next few years in the country and the same is true for other low- and middle-income countries (LMICs). Several energy monitoring systems (EnMS) have been proposed, but they face limitations in terms of cost, ease of application, lack of universal installation capability, customization, and data security. The research below focused on the development of an economical, secure, and customizable real-time EnMS. The proposed EnMS comprises low-cost hardware for gathering energy data with universal compatibility, a secured communication module for real-time data transmission, and a dashboard application for visualization of real-time energy consumption in a user-preferred manner, making the information easily accessible and actionable. The experimental results and analysis revealed that approximately 40% cost savings in EnMS development could be achieved compared to other commercially available EnMSs. The performance of the EnMS hardware was evaluated and validated through rigorous on-site experiments. The front-end of the EnMS was assessed through surveys and was found to be interactive and user-friendly for the target clients. The developed EnMS architecture was found to be an economical end-product and an appropriate approach for small and medium clients such as residential, institutional, commercial, and industrial consumers, all on one platform.