Skip to main content
Ko Lab

Research

Saghar works at the intersection of digital electronic engineering and artificial intelligence, building vision systems for practical settings such as smart transportation and automotive applications. Her focus is on closing the gap between model development and deployment through compact, task-specific designs that balance recognition accuracy against computational cost.

Current project

A coarse-to-fine framework for real-time HAZMAT placard recognition

Her thesis develops an efficient, application-specific computer vision framework for recognising hazardous material placards in complex real-world scenes. The pipeline adapts and optimizes deep learning models for multi-class object detection, classification, semantic information extraction and text and number recognition, progressively converting a raw scene into structured, machine-readable hazard information.

  • Computation is allocated according to the difficulty of each stage rather than uniformly across the pipeline
  • Three cooperating models: placard localization and classification, text and symbol localization, then OCR for UN/NA numbers and hazard text
  • Targets GPU, edge and real-time deployment on constrained hardware
Coarse-to-fine HAZMAT recognition pipeline: a first model localizes and classifies placards on a trailer, a crop-and-resize step feeds a second model that localizes class IDs, text and UN/NA numbers, and a third OCR model reads them, producing a structured record of hazard category, class ID, text and UN/NA number.

Stepwise optimization of the detector reduces the model to 1.0 M parameters and 3.8 GFLOPs while reaching 96.2% mAP@50 and 98 FPS on an NVIDIA Jetson Orin NX, using channel shrinkage, depthwise separable convolution, efficient channel attention and knowledge distillation on a long-tailed HAZMAT dataset.

Four-stage overview of the optimization framework: the long-tailed HAZMAT detection problem and its challenges, a YOLOv9-t baseline, the proposed stepwise optimization combining channel shrinkage, depthwise separable convolution and efficient channel attention with knowledge distillation, and the resulting real-time performance on an NVIDIA Jetson Orin NX.

Research areas

Multi-class object detection and classification

Efficient deep learning models that detect and classify multiple visual categories in cluttered scenes, with an emphasis on compact architectures, model optimization and deployment-oriented performance.

Text and structured information recognition

Extracting textual and numerical information from images through OCR, text recognition and specialised recognition models, so that visual data becomes structured and machine-readable.

Application-specific model design and optimization

Designing models around the characteristics and computational budget of an individual task, using task-specific processing paths, model compression, knowledge distillation, GPU optimization and efficient inference for edge and real-time systems.

Contributions to the lab

  • CADASIL medical imaging research on white matter hyperintensities, through dataset preparation, preprocessing and quality control
  • The CSM project, through literature review and investigation of sparse data formats and related computational approaches

Education

MSc, Electrical and Computer Engineering

University of Saskatchewan, 2024-present. Application-specific deep learning for smart transportation and real-world computer vision, covering multi-class object detection, classification, OCR and text and number recognition, model compression and efficient real-time inference for edge deployment. Supervised by Prof. Seok-Bum Ko.

BSc, Digital Electronic Engineering

K. N. Toosi University of Technology, Iran, 2023. Undergraduate work included drowsiness detection and IoT-based systems.