Digital Data in Biodiversity Research Conference

Detecting and cataloging fish specimens in datasets of digitized museum specimens is essential for scientific research, biodiversity studies, and conservation efforts. In this study, we use a machine learning approach utilizing YOLOv8, a state of art object detection algorithm, for detection of fish specimens within the collection images. Our purpose is to solve problems in datasets where there are multiple specimens and/or specimens have different orientations.
Our method addresses the challenges posed by varying fish sizes, orientations, and positions within the dataset. Leveraging YOLOv8's capability to detect objects with different orientations, we annotated the museum images and trained a model showing diversity among the museums in Great Lakes Invasives Network (GLIN) dataset and compared the results with manually observed number of specimen values from FishAIR (http://www.fishair.org) database. We also developed a model for capturing non-specimen objects such as labels, scale bars, and color bars.
Our model resulted in 99.4% accuracy finding fish specimens with the parameters 75 epochs and 0.5 score threshold. This study presents a practical and efficient solution for automating the detection of multiple/oriented fish specimens as well as single specimens in museum image data, aiding researchers, curators, and conservationists in their endeavors to document and preserve biodiversity effectively.
This research is supported by: NSF-HDR #2118240 (Imageomics)

Co-author: Yasin Bakis ,Tulane University; Xiaojun Wang, Tulane University; Henry Bart , Tulane University