Research profile J (Jiwoo) Ryu, MSc

Jiwoo Ryu, Msc

Project: Low-complexity, parallel and distributed algorithms to detect and classify objects in large infrared, hyper-spectral and 3D sensor images

This project (Bernoulli, Kapteyn) will develop low-complexity, parallel and distributed algorithms that robustly detect and classify objects in large infrared, hyper-spectral and 3D sensor images. The methodology will be based on computer vision and machine learning techniques, such as trainable morphological image processing filters which can be efficiently implemented using sequential and parallel implementations based on the max-tree or related alpha-tree data structure. We intend to explore two ways of combining these morphological scale-space data structures with machine learning, i.e. 1) using machine learning to classify nodes in the trees based on so-called vector-attribute filtering, and, 2) feeding key-point features detected by analysis of these scale spaces to deep convolutional networks, after automatic rescaling and rotation to a standard scale and orientation. This would allow scale-invariant analysis of huge images using deep learning, without excessive compute power requirements. The methods will be applied to several use cases: 1) the detection of buildings at many different types of resolution from remote-sensing data, as needed in the Global Human Settlement Layer project which aims to improve understanding of urbanisation, to assist urban planning, and 2) to support disaster relief, 3) for detection and analysis of objects in large astronomical surveys. This subproject will profit from existing collaborations of JBI with the Joint Research Centre (JRC) in Ispra, Italy. Potentially, a fourth use case on large electron micrographs obtained from the UMCG could be included.

Keywords: Scale spaces, connected filters, machine learning, remote sensing, astronomical surveys.