A Convolutional neural network (CNN) based approach to recognize medicinal plants by analyzing plant leaf

Abstract

country. Most of these plants can be used for therapeutic purposes like when in ancient days, our ancestors used them vastly and considered them to be one of the most highly efficient medical systems in the world at that time. However, with the dawn of modern medicine, indigenous medicine has been decreasing in usage due to factors such as the lack of knowledge about medical plants, the desire for fast recovery, and the reducing interest in traditional treatments due to their smells and appearances. Out of the above, the lack of knowledge about medical plants has been identified as the most contributing factor that demotivates the general public from using traditional medicine. Hence it is evident that a reliable and easy-to-use application to identify and analyse medical plants would be a timely solution to increase the use of traditional medicine in society today. The main objective of this research was to review the features used for leaf recognition, evaluate existing leaf-based medicinal recognition systems, and design a system that would address the loopholes in the available solutions. As such, the researcher carried out a comprehensive literature survey and reviewed existing classification methodologies like Support Vector Machine, Principle Component Analysis, Probabilistic Neural Network and Conventional Neural Network to assess what the best methodology for the above task would be. Due to its feature-extraction capability and high levels of accuracy, Conventional Neural Network was selected as the best approach for this study. Based on the selected method, the researcher designed and developed a feasible application that is capable of finding medical plants by the features of its leaf and medical values. Once the system was built, the researcher distributed surveys and conducted interviews in order to critically test and evaluate the application among industrial experts as well as general users. An overall recognition rate of 85% was recorded by the system, which was well-appreciated by the experts. However, recommendations like probable scope expansions and the need for higher response times were also suggested in the feedback received