Automated Detection Of Herbarium Specimens Via Transfer Learning In Convolutional Neural Networks

There are thousands of herbaria (collections of dried and mounted plants) all over the world, containing millions of specimens which have yet to be digitized and made available to online research communities. Recent global transcription efforts have utilized crowd-sourced volunteers to perform data entry, especially in areas where optical character recognition continues to fail. The relatively new process of transfer learning in artificial neural networks has shown promise in reducing training complexity in difficult image classification problems, despite notable differences in target tasks and domains. Within this work, the technique of transfer learning is applied to the digital specimen collection of the I.W. Carpenter Jr. Herbarium housed at Appalachian State University, in an effort to assess its feasibility. It is shown that within the confines of the ASU herbarium, the technique of transfer learning combined with modern neural networks can effectively classify specimen images to the point where volunteer-based transcriptions of certain fields may no longer be necessary.