Development of a deep learning-based insect recognition system : a comparative analysis of different YOLO models

Abstract

Insect populations play pivotal roles in both agricultural ecosystems and biodiversity monitoring, yet traditional identification methods remain labor-intensive and error-prone. This thesis presents a deep learning–based insect recognition system leveraging three variants of YOLOv8 object detector—nano (v8n), small (v8s), and large (v8l)—to balance detection speed, model size, and accuracy. In this research, an image dataset comprising diverse insect taxa was curated, and various models were subsequently trained. A systematic comparison of their computational speed, model size, and accuracy was then conducted, employing metrics such as precision, recall, F1-score, and mean Average Precision (mAP) for performance evaluation. The resulting models were integrated into a user‑friendly GUI that accepts offline images and displays bounding boxes, species labels, confidence scores, and inference times.

Experimental evaluation shows that the large variant achieves the highest accuracy, while the small variant offers the most balanced trade‑off between computational cost and performance. Distinctive taxa attain high detection rates across all models, whereas under‑represented or visually subtle groups remain challenging. This work lays the foundation for scalable, real‑world insect monitoring and outlines future directions for robustness enhancement and edge deployment.