Low complexity and accurate Machine learning model for waterborne pathogen classification using only three handcrafted features from optofluidic images

Abstract

Pathogenic microorganism can cause human diseases and even death, so waterborne pathogen detection is essential for water quality monitoring. Conventional microorganism detection/classification through biochemical approaches is expensive, time-consuming, labour-intensive and leaves chemical pollutions. Recently, many image processing algorithms are developed for cheaper, faster, more automatic, and more environmentally friendly microorganism classification. However, most existing classification algorithms are designed for microscopic images and focus only on the classification performance, not the computational cost. Alternatively, the optofluidic systems provide a smaller and cheaper way to capture microorganism images than conventional microscopes. Thus, in this paper, an image classification algorithm based on optofluidic images is proposed for portable, real-time water quality monitoring applications. With a comprehensive investigation on different categories of features and careful analysis on our optofluidic images, 2 morphological features were selected to classify the waterborne pathogens with a proposed feature, EHHOG (entropy of histogram of HOG) (HOG: histogram of oriented gradients). The proposed algorithm can achieve higher or similar classification accuracy as existing methods, while the computational cost is much lower with only 3 features.