Fungus Detection Using Computer Vision and Machine Learning Techniques

Fungus is extremely disreputable and dangerous for food, human health and archives because it causes food loss, various life threatening diseases to human and destroy important documents. Thousands of different fungus species exist in the world and their spores always present indoor and outdoor environments. Its sign and symptom are non-specific in medical science for extremely large areas and containers, which poses severe threat to human health, food and archives. Numerous traditional techniques were applied to meet the challenge of early detection of fungus but all are costly, laborious, time-consuming and required skilled staff. This revolutionary era emphasizes the need for novel, simple, automatic fungal detection system to control the devastation caused by fungal species. In this research, we develop a fungus detection system and algorithms to automatically detect fungus. Three computer vision based techniques were developed for the detection of fungus spores. One of them used HOG based features and achieved convincing results. Other technique consisted of fusion of Fourier transform and SIFT features to achieve promising results. Third method based on superpixel and handcrafted features. The results of all these techniques encourage for the possibility of early detection of fungus spores from dirt particles. The other main objective of this research was to develop a CNN based approach for the detection and classification of different types of fungus spores. However, a large amount of data is an essential prerequisite for its effective application. In pursuing this idea, we developed a new novel fungus dataset of its kind, with the goal of advancing the state-of-the-art in fungus classification by placing the question of fungus detection. This is achieved by gathering various images of complex fungal spores by extracting samples from contaminated fruits, archives and lab incubated fungus colonies. These images primarily consisted of five different types of fungus spores and dirt. The fungus detection system was utilized to obtain these images. Which were further annotated to mark fungal spores as a region of interest using specially designed graphical user interface. As a result, 40,800 labeled images were used to develop fungus dataset to aid in precise fungus detection and classification. A CNN architecture was designed and it showed the promising result with an accuracy of 94.8%. The obtained results proved the possibility of early detection and classification of several types of fungus spores using CNN and could estimate all possible threats due to fungus.