Training of low-altitude city image enhancement models for semantic segmentation
||Training of low-altitude city image enhancement models for semantic segmentation|
||Alexander Bock <email@example.com>|
||2023-10-19 – 2024-02-01|
The project aims to investigate the improvement effects of enhanced, i.e. upscaled, images on image classification models to be used in, among other fields, urban visualization. The models in question are used to classify images taken from airplanes. The goal is to train different models using image datasets from different seasons of the year as the individual variety of image features between seasons reduces the accuracy of the image upscaling model.
In order to investigate this, image enhancement models that are trained on large sets of drone images acquired throughout the year are required. Due to the necessary size of the datasets, the training of the enhancement models currently require the use of supercomputer infrastructure and is thus uniquely suited for the Berzelius cluster.
The project will benefit the research community as it will help the understanding of how image enhancement in the form of heighten resolution will benefit image classification in other fields than airborne image acquisition.
In the long run, better image classifications of drone, airplane and satellite images are in turn beneficial and needed in societal areas such as city planning, city maintenance, environmental monitoring and potential disaster management.