A lightweight detection model for fruit flies based on KSGM-OMN

In the agricultural sector, the fruit and vegetable industry plays a pivotal role.Fruit fly larvae invade fruits and vegetables, causing them to rot and seriously affecting the yield and quality of these crops. Traditional pest control methods relying on pesticides not only trigger ecological environmental pollution but also pose potential risks to human health. Therefore, the development of accurate and efficient fruit fly detection technologies has become a critical need in agricultural research and practice. To address the demands for accuracy and efficiency in fruit fly detection, this study aims to develop a lightweight deep learning-based model for detecting small fruit fly targets, overcoming the limitations of traditional detection methods. To further enhance detection efficiency, this research proposes a lightweight fruit fly detection model, KSGM-OMN (Key Sample Guided Mapping - Object Mapping Network). Herein, the Object Mapping Network dynamically selects high-quality samples through an ingenious center assignment strategy, effectively reducing hardware resource consumption and accelerating the model training and inference processes. The Focal Loss function successfully addresses the issue of positive-negative sample imbalance, while the Adaptive Gating Target (AGT) inference mechanism accurately filters grids containing targets, greatly improving detection efficiency. Experiments were conducted on the self-constructed Drosophila-Four dataset and the public IP102 dataset. The results demonstrate that, building upon the KSGM-YOLO model, the KSGM-OMN model achieves a detection speed of 40 FPS on the same equipment, with an accuracy of 72.1%, outperforming current mainstream detection models with low parameter counts.