Object Detection of Small Insects in Time-Lapse Camera Recordings

Area of Science:

• Ecology and Environmental Science
• Computer Science and Artificial Intelligence

Background:

• Insects are vital pollinators for ecosystem health and global food security.
• Declining insect populations worldwide necessitate advanced monitoring techniques.
• Current methods struggle with detecting small insects in complex natural environments using time-lapse imagery.

Purpose of the Study:

• To develop and validate an automated method for detecting insects in time-lapse RGB images.
• To create a comprehensive dataset of annotated insect images for training and testing detection models.
• To improve the accuracy of insect detection compared to existing deep learning approaches.

Main Methods:

• A novel two-step approach combining motion-informed image enhancement with convolutional neural network (CNN) object detection.
• Preprocessing of time-lapse images to highlight insects using motion and color cues.
• Utilizing and adapting You Only Look Once (YOLO) and Faster Region-based CNN (Faster R-CNN) object detectors.

Main Results:

• A dataset of 107,387 annotated images featuring primarily honeybees, with 9423 annotated insects.
• Motion-informed enhancement significantly improved insect detection performance.
• The YOLO detector's micro F1-score increased from 0.49 to 0.71, and Faster R-CNN's from 0.32 to 0.56.

Conclusions:

• The proposed motion-informed enhancement technique effectively improves insect detection in time-lapse imagery.
• The developed dataset and method represent a significant advancement in automated monitoring of flying insects.
• This work supports ecological research and conservation efforts by enabling more efficient insect population studies.