Ultra high-resolution segmentation of biological images using reduced-memory convolutional neural networks with inter-tile communication

Benson, Ezenwoko (2022) Ultra high-resolution segmentation of biological images using reduced-memory convolutional neural networks with inter-tile communication. PhD thesis, University of Nottingham.

[img] PDF (Thesis - as examined) - Repository staff only - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution Non-commercial.
Download (133MB)


Advances in imaging technology continue to outpace the innovations in computing hardware. For some tasks such as object classification this gap presents almost no problem, given that the results are largely unaffected if some pixel data is missing. This task can produce accurate results by shrinking the image since not all pixel information is required. As such, currently available hardware can be used to complete this task. Tasks like segmentation produce dense per-pixel classifications and benefit greatly from processing all of the pixels in an image, unchanged. In this task only a small amount of image shrinking can take place before the results are affected. Small images can be segmented however, the computational cost becomes prohibitive when applying the same methods to high resolution images. This hasn’t been adequately addressed by current research because of the focus on performance on low resolution benchmark datasets. This thesis presents a novel solution which builds upon the commonly used technique of image tiling. Image tiling is where an image is divided into smaller sub-images called tiles. Processing tiles reduces the memory burden therefore, allowing for the segmentation of larger images. Segmenting a single tile only allows the method to view the context within a tile, which results in inaccurate segmentations when a wider context is required. In this work, context is shared between tiles using global filters. This allows corrections to be made to predictions based on the wider context of the image. The method is shown to be competitive on low resolution datasets against a baseline. It is also shown to outperform high resolution segmentation techniques on high resolution datasets.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Pound, Michael
Pridmore, Tony
French, Andrew
Keywords: deep learning, high-resolution images, convolutional neural networks, global context
Subjects: Q Science > Q Science (General)
Q Science > QA Mathematics > QA 75 Electronic computers. Computer science
Faculties/Schools: UK Campuses > Faculty of Science > School of Computer Science
Item ID: 68434
Depositing User: Benson, Ezenwoko
Date Deposited: 02 Aug 2022 04:40
Last Modified: 02 Aug 2022 04:40
URI: https://eprints.nottingham.ac.uk/id/eprint/68434

Actions (Archive Staff Only)

Edit View Edit View