Year: 2026
Type: article
Abstract: Agricultural pests pose a persistent threat to crop yield and quality, making accurate and efficient detection essential for modern crop protection. Although trap-based systems are widely used for automatic pest monitoring, pest detection in captured images remains challenging due to tiny pest bodies, clustered distributions, diverse capture states, and inter-class similarity. To address these challenges, we propose PestFG-RTDETR, a detector with three key components: frequency-guided adaptive feature interaction, dynamic upsampling, and a lightweight backbone. We also introduce a biologically inspired Copy-Paste augmentation strategy to enrich minority classes and generate realistic dense pest layouts on the dataset. When evaluated at an input resolution of 416 × 416 pixels, PestFG-RTDETR achieves an [email protected] of 61.64% and an [email protected]:0.95 of 37.27%, outperforming RT-DETR-R18 by 4.67 and 6.08 percentage points, respectively, while using 1.55M fewer parameters. Extensive comparisons with representative two-stage, one-stage, and transformer-based detectors show that PestFG-RTDETR achieves competitive performance, indicating the effectiveness of the proposed design. Independent training and evaluation on a self-collected rice pest dataset further indicate that the proposed architecture can be applied to another trap-based pest detection scenario.
Source: Frontiers in Computer Science
Authors: Meng Zhang, Xianliang Jiang, Zijian Le, Guang Jin
Institution: Ningbo University
Language: English
Topic: Smart Agriculture and AI
Subfield: Plant Science
Domain: Life Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $1,418
Award: 融合多源感知的农业虫害监测预警系统研究
OpenAlex ID: W7224869569