Artificial intelligence remains massively underutilized by entomologists, despite its well known potential as a scalable, accessible solution to the problem accurately identifying the millions of extant insect species. Recent advances in digital photography, computer vision approaches, and user-friendly libraries, have minimized the technical barriers to widespread adoption of automated insect identification applications. The main obstacle now appears to be a lack of familiarity with computer vision among the relevant domain experts: taxonomists and other collections-based researchers. Here, I present an easily adaptable system for training and deploying a computer vision model using museum specimens. I use the parasitoid wasp family Ichneumonidae as an example by developing an automated identification app based on a frontal view of the head. With approximately 6,000 non-focus-stacked images of ichneumonid heads, the model achieved a 95% accuracy identifying 34 ichneumonids subfamilies and selected tribes and genera. The model was then deployed atichsofna.org using a minimalistic client-side approach with inference occuring within the user’s web browser. Finally, I discuss major considerations for adapting the system to other taxa including: efficiently generating images while minimizing bias, the use for iterative testing to maximize the number of included taxa and identify weaknesses in the model, and why taxonomists themselves need to lead the development of computer vision models.
.png)
