Deep Convolutional Neural Network based Classification of Microscopic Images of Environmental Microorganism

Abstract

Environmental microorganisms (EMs) plays a critical role in the development and sustainability of human civilization. Detailed study and analysis of EM will be important while carrying out research in areas like waste management, agriculture, green technology, etc. Currently, the classification of EM using microscopic images is manually intensive and requires domain experts. Hence, in this domain, there is a scarcity of existing standard datasets for carrying out useful research. EMDS-6 is one of the standard EM microscopic image data set consisting of 21 types of EMs. However, extracting and analyzing important features from a small EMDS-6 dataset using data-intensive Deep Learning (DL) approaches is challenging. In this work, we compared various Deep Convolutional Neural Network (DCNN) models along with a data augmentation strategy for EMDS-6 dataset classification with good accuracy. After extensive experimentation and detailed ablation study, we found that MobilenetV2 pre-trained model with three dense layers; with each dense layer using 'SELU' activation function, provides the highest training and testing accuracy with less number of parameters.