Mathematical Morphology in Machine Learning

  • Cleopatra M. Y. Hammad

Student thesis: Master's Thesis

Abstract

Mathematical Morphology (MM) is a well-known field in visual computing that is concerned with geometrical structures. In this research, MM is used in Machine Learning (ML) by using the MM-based classifier, named Morphological Dilator Classifier (MDC). As the name implies, MDC is dependent on the 'dilation' morphological operator. The main concept of MDC model is to create a feature-space by dilating around the training instances – which are used as the initial seeds. In other words, labeling the space around the initial seeds according to classes, while respecting the boundary rules of the MDC algorithm. Once the feature-space is fully explored after several iterations, unlabeled instances are used to investigate the constructed model's performance. Datasets of low- to medium-dimensions and binary or multi-class problems, are used to test the MDC against state-of-the-art classifiers. The implementation of state-of-the-art classifiers is performed using MATLAB's Classification Learner Application and MATLAB's Neural Network Toolbox. Performance comparison between the MDC and other classifiers is obtained using the Receiver Operating Characteristic (ROC) Curve, Accuracy, Confusion Matrix, Sensitivity and Specificity as performance metrics. MDC showed significant increases of 11.93% and 10.04% in the Skin Segmentation and Iris datasets, respectively, compared to the mean accuracy of the state-of-the-art classifiers. However, it performed very poorly in the case of the imbalanced Hayes-Roth and Balance datasets. The poor performance of the MDC could be rectified using both PCA techniques and class imbalance methods (e.g., SMOTE). Moreover, the application of the proposed set of classifiers in high (>1000) dimensional datasets requires further investigation by means of efficient representation of feature space and applying mathematical morphology operators directly on the associated data structures. In summary, the MM-based framework has shown some promise in classification tasks, and additional investigations are necessary to further improve its efficiency and performance.
Date of AwardMar 2019
Original languageAmerican English

Keywords

  • Mathematical Morphology (MM)
  • Machine Learning (ML)
  • Supervised Learning
  • Unsupervised Learning
  • Morphological Dilation.

Cite this

'