Jordan Shapes for Deep Learning

The proposed project aims to develop a systematic approach for improving deep-learning-based computer vision systems by augmenting the local pixel data with the global shape data (more specifically, Jordan curves) and by adjusting system architectures to accommodate the augmented input. Three canonical computer vision problems will be investigated in this project. They are respectively image dehazing, alpha-matting, and face detection. The potential roles of Jordan curves in these applications will be examined. The research results will provide an in-depth analysis on how shape data may impact deep learning training, inference and transparency and suggest a general guideline on how to effectively utilize shape data in deep learning systems.

Chenxiao Niu
Faculty Supervisor: 
Jun Chen
Partner University: