| Summary: | Waterborne pathogens are microorganisms in the water that can cause infections, diseases and
even deaths. Classification of such microorganisms is essential for water quality monitoring
which is important for human health and safety. The conventional biochemical approaches
are standard, accurate but have high requirements on monetary, time, and human resources,
and usually leave chemical pollutions. With the development of machine learning techniques,
algorithms were established for cheaper and simpler microorganism image classification.
This PhD study has investigated and developed portable and accurate classification algorithms
that are more feasible for portable, real-time applications based on images captured by
an optofluidic system, which is smaller and cheaper than microscopes. Most existing microorganism
classification algorithms are designed for microscopic images so do not work well on
our optofluidic images. Besides, the existing algorithms are mainly focused on classification
accuracy but not on the computational cost. Thus, this PhD study aims to develop image processing
algorithms for waterborne pathogen classification that have the potential for portable,
real-time water quality monitoring applications.
This PhD study has made three contributions. The first contribution is the proposed
EHHOG (entropy of histogram of HOG) (HOG: histogram of oriented gradients) method,
combining some existing morphological features extracted from binarized images and the
proposed feature EHHOG extracted from greyscale images, for waterborne pathogen classification.
The second contribution is proposing the Multi-Region Binarization method to capture
the information missed by the existing binarization methods and developing the Binarized-
Greyscale-Hybrid Algorithm with Multi-Region Binarization (BiGHAM) method that improved
the classification accuracy from the EHHOG method. The third contribution is the proposed
improved-BiGHAM (iBiGHAM) method with a new circle-based region partitioning
for Multi-Region Binarization and a new way to apply the mutual-information-based feature
selection. The iBiGHAM method further increased the classification accuracy with simpler
trial procedures than the BiGHAM method. It could achieve similar or higher accuracy than
the existing algorithms including the deep-learning-based approach, while the computational
complexity was much lower. Moreover, these proposed algorithms were adapted and applied
to images in an open-sourced dataset, also achieving better classification performance than
the existing algorithms.
|
|---|