An Empirical Evaluation of Lightweight YOLO Architectures for Plant Pest Detection

Abstract

Automated pest detection plays a critical role in supporting agricultural productivity by enabling accurate and efficient recognition of pest species in field conditions. This study presents a comparative evaluation and ablation analysis of four lightweight object detection models, namely YOLOv5n, YOLOv8n, YOLOv11n, and YOLOv12n, for plant pest detection tasks. All models were trained and tested on the IP102 dataset, a large-scale benchmark comprising 102 pest categories captured across diverse field environments. Each model was evaluated under four optimization scenarios: AdamW optimizer, Mixup and Mosaic augmentation, a combination of both strategies, and Test-Time Augmentation (TTA). Performance was measured using precision, recall, F1-score, mAP@0.5, and mAP@0.5:0.95 to assess both detection accuracy and bounding box localization quality. The results demonstrate that YOLOv12n consistently achieves superior performance across the majority of evaluation metrics. Under the TTA scenario, YOLOv12n attained a precision of 0.590, recall of 0.698, F1-score of 0.639, mAP@0.5 of 0.689, and mAP@0.5:0.95 of 0.444, representing the highest scores among all evaluated configurations. TTA proved to be the most consistent optimization strategy, delivering stable improvements across all model architectures. In contrast, combining AdamW with data augmentation degraded performance across all models, likely due to over-regularization effects in lightweight network designs. These findings highlight that inference-level strategies can be more beneficial than training-level augmentation for compact detection models. This study provides practical insights for selecting and optimizing lightweight detectors in real-world agricultural deployment scenarios.