Efficient Representation Learning for Optical Image Analysis
Liu, Quan
0000-0002-8609-3556
:
2024-07-12
Abstract
This thesis addresses the challenges of efficient representation learning across three crucial categories of optical images: microscopic, pathology, and meta-optic images.
Microscopic images, characterized by dense dynamic objects, often suffer from resource-intensive object annotations. To tackle this, we introduce a novel deep learning-based unsupervised sub-cellular microvilli segmentation method and propose an annotation-free video analysis paradigm.
In the realm of medical optical images, we propose the SimTriplet approach, integrating GPU memory-efficient techniques with self-supervised learning to enhance computational efficiency and feature extraction.
Moreover, we tackle the escalating computational demands by introducing a large convolution kernel design for LMNN models, reducing computational latency and energy consumption. We validate our approach through physical meta-material fabrication.