Virus Identification From Digital Images Using Deep Learning
Given the electron microscopy images, virus recognition using deep learning approaches from digital images is critical at present, because virus identification by human virology experts is slow and time-consuming, this research project aims to develop a deep learning-based method for automatic virus detection. There are four virus species in this thesis, they are SARS, MERS, HIV, and COVID-19. This study is based on classification and bounding box regression. In this thesis, we firstly examine virus morphological characteristics and propose a novel loss function which targets to reflect the viruses on the given electron micrograph. In this project, we take into account the attention mechanism, virus images are processed in advance to be trained for classification and localization. In order to make the best estimation of bounding boxes and classification for a virus, we test five deep learning networks: R-CNN, Fast R-CNN, Faster R-CNN, YOLO, and SSD, based on the prior knowledge of virus electron microscopy. Additionally, in this project, we discuss the deep learning training problems and illustrate the evaluation approaches. The conclusion reveals SSD and Faster R-CNN outperform in the virus detection from digital images.