Hui Yu

Developed YOLOv9-based object detection pipelines for insect detection in camera-trap images, addressing scale variation and cluttered backgrounds. Integrated sliding-window (SAHI) inference to improve small object detection and analyzed trade-offs between accuracy and inference cost. Applied semi-supervised learning with paste-in augmentation to handle limited annotations and domain shift. Published results at CamTrap Workshop and submitted extended work to IET journal.

Developed a C++ passenger counting and seat occupancy system on Linux edge devices for real-time monitoring in public transport. Deployed Dockerized applications (Alpine Linux, OpenRC) for reliable integration and auto-recovery. Built APIs for HTTP image retrieval and RTSP streaming with MQTT JSON messaging. Set up W&B benchmarking pipelines and optimized models for Hailo/NCNN, improving accuracy by 20%. Implemented Kafka-based messaging for scalable video analytics.

- Implemented a ViT-based self-supervised contrastive learning approach with DINO, enhancing feature extraction robustness for unstained white blood cell images - Developed a fully transformer-based deep learning pipeline for patient-wise classification, optimizing end-to-end performance in medical applications - Standardized the annotation process and established refined guidelines to improve consistency and accuracy in data labeling

- Proposed a patient-wise classification strategy using transformers, achieving a 10% accuracy improvement compared to conventional CNN-based approaches - Implemented a deep learning classification pipeline tailored for unstained white blood cell images - Developed GAN-based data augmentation techniques for enhancing datasets with pathological and immature cell images, supporting improved model generalization

- Grade: 1.4 - Master’s Thesis: Unstained White Blood Cells Classification Using Deep Learning Strategy - Relevant Courses: Pattern Recognition, Computer Graphics, Applied Visualization, Deep Learning, Computer Vision, VR Technology

- Grade: 2.3 - Bachelor’s Thesis: Simulation of Beidou Satellite Navigation System with Unity3D - Relevant Courses: Object-Oriented Programming C++, Data Structure, Software Development, Image Processing, Artificial Intelligence, Machine Learning