Preventing exotic pet beetle invasion with an improved lightweight and efficient pest detection model deployed on...

The illegal smuggling of exotic pet beetles presents a growing threat to global ecosystems. Customs authorities play a critical role in preventing biological invasions, yet current identification methods rely heavily on expert knowledge and time-consuming laboratory analysis, which limits rapid responses at ports of entry. To address this issue, we propose EPB-YOLO-PD, a lightweight, mobile-deployable detection model for real-time recognition of exotic pet beetles. The source code is available at https://github.com/bihaojie/EPB-YOLO-PD.

EPB-YOLO-PD incorporates three originally designed components - the Feature Aggregation and Mixing Network (FAMNet), Multi-Scale Efficient Lightweight Optimization Network (MELON), and Partial Multi-Head Self-Attention Residual Block (C4PMS) - along with an improved detection head (CAHead), and a newly introduced loss function (Slide Loss). Structural pruning and knowledge distillation are applied to reduce model size and improve inference speed. When tested on a custom dataset of 13 intercepted species, the model achieved detection accuracies between 93.3% and 99.3%. Compared to the YOLOv11n baseline, EPB-YOLO-PD demonstrated a 2.0% increase in mAP0.5 (97.3%), a 74.04% reduction in model size (1.35 MB), and a 65.08% decrease in computational complexity (2.2 GFLOPs). The PetBeetle Finder app, based on this model, runs at over 25 frames per second (FPS) on a Huawei Mate 40 smartphone.

EPB-YOLO-PD offers an effective solution for real-time detection of exotic pet beetles at customs checkpoints. It enables rapid and accurate classification, effectively handling challenging scenarios such as incomplete morphological features and visually confusing backgrounds, and provides a replicable framework for intercepting other invasive species. © 2025 Society of Chemical Industry.