Exploring Hybrid Morphological Neural Networks: A Comparative Study on CNNs

Huese, Reinier (2025) Exploring Hybrid Morphological Neural Networks: A Comparative Study on CNNs. Bachelor's Thesis, Computing Science.

Preview	Text Exploring-Hybrid--Morphological-Neural--Networks-A-Comparative--Study-on-CNNs.pdf Download (2MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (179kB)

Abstract

Several studies have shown an increased performance when forming a combination of CNNs and MNNs into a Morphological Convolutional Neural Network (MCNN). We theorize that increased performance with the same training data volume can be translated to the same performance with less training data. This would lower the requirements of getting a custom neural network, as less data is required for the purpose of training. In our experiments, the MCNN has shown a significantly decreased accuracy in most cases with Gaussian noise and spectrally uncorrelated salt and pepper noise. For spectrally correlated salt and pepper noise, the MCNN achieves significantly higher performance compared to the CNN in all cases. Both models show slight improvement when training on the Indian Pines (IP) dataset, prior to training and testing on the University of Pavia (UP) set, however the effect is insignificant in most cases. The retraining does result in the MCNN achieving significantly higher accuracies when compared to the CNN for certain low training percentages, where there was no significant difference before. Additionally, we have found that the MCNN is able to achieve significantly higher accuracies compared to the CNN when training on 1% or more of the IP dataset, and 2% or more when trained on the UP dataset. Another observation is that to achieve higher performance than the MCNN that has been trained with 5% of either dataset, the CNN needs to be trained on 15% of that dataset.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Wilkinson, M.H.F. and Bunte, K.
Degree programme:	Computing Science
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	18 Jul 2025 06:30
Last Modified:	18 Jul 2025 06:30
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36377

Actions (login required)

View Item