Development of Chrome Extension for real-time YOLO-based object detection using TensorFlow.js

Abstract

This thesis investigates how deep learning models can be efficiently deployed in client-side web browser environments through a practical implementation of a real-time object detection system. The project demonstrates this by converting and integrating a YOLO11n model into a Chrome Extension using TensorFlow.js. The extension performs all inference locally in the browser using WebGL acceleration, allowing for real-time object detection without sending data to external servers. Key features include keyword-based filtering, visual overlays, and adaptive performance management. The system was evaluated across different hardware configurations, and was compared with a related web-based application. The results support the feasibility of browser-based deep learning as a privacy-preserving alternative to server-side inference.