A Convoluted Neural Network for Object Detection of Common Flower-Visiting Insects

Abstract

An important aspect of pollinator ecology is monitoring and identifying the kinds of animal visitors to flowers. Conventional methods often use video recording or camera traps to collect data and human analysts for manual review. However, because of the large volumes of data collected and partiality in human reviewers, these methods are costly in terms of time and resources, can be prone to human error, and require taxonomic expertise. Computer vision (CV) is a branch of deep learning that has demonstrated promise as a lower-cost and rapid alternative to conventional methods for pollinator monitoring. Object detectors are a kind of deep learning network which use CV to identify and localize multiple different objects within an image. An object detector may be placed on a camera in the field to process in real time video and record different visitors to flowers, thus cutting out time for manual review. For this project, an object detector will be trained on 10,000 images from the iNaturalist research-grade database sourced from GBIF. Images are distributed across eight classes (Bees, Lepidoptera, Flies, Beetles, Wasps, Bugs, Ants, & Spiders) representing different common taxa observed at flowers and weighted according to frequency of visit and importance. Bees and Lepidoptera are weighted heaviest and spiders least. Several different architectures of the YOLO object detector infrastructure will be tested. Preliminary results show that this architecture will be useful for monitoring of pollinators in the field.