University of North Dakota
Adaptblur: Adaptive Linear Filter For Enhanced Deep Learning Classification Performance
Abstract
dc:description.abstractImage preprocessing is crucial in deep learning models for high performance and accuracy. This study implements AdaptBlur, a dynamic linear filter developed from a second-order partial differential equation using finite difference approximations, which enhances image quality while preserving the image structure with dynamic parameters that are optimized through the Nelder-Mead optimization, minimizing mean square error to improve the effectiveness of the filter. Moreover, the study evaluates the impact of the results on model performance on deep learning classification using publicly available datasets. Experimentally, image classifiers trained on preprocessed data with the AdaptBlur filter perform much better than those trained without filtering or with filtering using the conventional Gaussian filter, as this filter gives nearly a 10% increase in classification accuracy.
Degree
thesis:*- Name thesis:degree_name
- Master of Science (MS)
- Level thesis:degree_level
- Thesis
- Discipline thesis:degree_discipline
- Mathematics
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Thondilege, Ganesha
- Contributors dc:contributor
-
- Bryce Christopherson
Subjects
dc:subject × 1Identifiers
dc:identifier.*- Repository record dc:identifier
- https://commons.und.edu/theses/7159