Work

RepGhostConv-Subpixel Fusion-Cascading Attention-Detection (RSCDet): A novel lightweight model for real-time aphid detection

Year: 2025

Type: article

Abstract: Efficient and cost-effective automated detection systems are essential for mitigating the impact of aphid infections on cereal crops. However, the fine-grained, small-scale, and densely clustered nature of aphid targets, combined with the computational demands and large size of existing detection models, poses challenges in achieving high detection accuracy on resource-constrained devices. To address these issues, we propose RSCDet (RepGhostConv-Subpixel Fusion-Cascading Attention-Detection), a lightweight deep learning model built upon the YOLOv7s architecture (You Only Look Once, version 7, small), which was selected after evaluating YOLO versions 5 through 10. RSCDet is designed to identify three morphologically similar aphid species, i.e., sugarcane aphid ( Melanaphis sacchari ), cereal aphid ( Macrosiphum avenae ), and bird cherry-oat aphid ( Rhopalosiphum padi ). RepGhostConv and Subpixel Fusion were used for network lightening. We implemented channel alignment and updated the detection head to reduce model size. Additionally, the SF-CasAM (Subpixel Fusion-based Cascading Attention Module) was developed to improve the recognition of visually similar targets, and the RepBi-PAN neck network was constructed to enhance the localization of small targets. RSCDet outperformed the baseline model, with higher detection accuracy (mean average precision, mAP of 95.8 %, increased by 2.4 %), faster processing speed (frames per second, FPS increased by 45.0 %), and lower computational complexity (parameters reduced by 93.5 %, FLOPs by 68.7 %, and model size by 91.5 %). It also demonstrated superior performance over other state-of-the-art models. RSCDet facilitates real-time, high-accuracy aphid detection on low-cost, portable embedded devices, enabling direct field deployment of early-warning systems and timely, actionable control of aphid outbreaks in agricultural settings.

Source: Computers and Electronics in Agriculture

DOI: 10.1016/j.compag.2025.110591

Authors: Weiyue Xu, Tianyi Wang, Tianyu Ji, Qiong Su, Wei Chen, Changying Ji