Lightweight object detection for drones - MSc project takes off

Congratulations to Zhifan Song, whose research which began as MSc project work has now been accepted for three conferences and a journal.

He's now joining Qualcomm for an internship, focusing on AI-driven processor optimisation and verification.

"Zhifan has a passion for real-time AI on edge devices, which led him to design lightweight object detection frameworks for drones, achieving impressive accuracy with reduced computational demands, resulting in multiple publications." Abdalrahman Abu Ebayyeh MSc project supervisor

Zhifan's project supervisor, EEE Teaching Fellow Abdalrahman Abu Ebayyeh, describes Zhifan's dedication and achievements as "remarkable."

The research focuses on vision AI for diverse, impactful applications, with an emphasis on lightweight and efficient AI designs. During Zhifan's MSc project at Imperial, he addressed the challenge of industrial defect detection, specifically targeting printed circuit boards and open-source microcontrollers. These devices often face quality issues in small-scale manufacturing, where variability in styles due to open-source designs makes it impractical to create comprehensive defect datasets.

Developing a real-time, accurate detection model under these constraints posed a unique challenge.

To overcome this, he devised innovative methods that bridged the gap between synthesised and real-world data using 3D modelling, data augmentation, and advanced AI block designs, producing robust generalisation that could not be achieved using existing object detection models.

Building on this foundation, his work expanded into edge AI, where he developed frameworks like EDNet, a small-target detection model optimised for drones, delivering superior accuracy with significantly reduced parameters compared to mainstream detectors like YOLO and lightweight models proposed in existing literature, enabling use in resource-constrained devices such as smartphones and IoT systems.

Additionally, he contributed to edge AI solutions for infrastructure inspection, introducing real-time UAV-based analysis that eliminates reliance on cloud computing while enhancing data privacy. Abdalrahman says "These projects have not only advanced edge AI capabilities but also set the stage for my specialisation in AI hardware accelerator design, which he pursued in Paris following his graduation from Imperial."