Authors

Abstract

When maize is afflicted by pests and diseases, it can lead to a drastic reduction in yield, causing significant economic losses to farmers. Therefore, accurate and efficient detection of maize pest species is crucial for targeted pest control during the management process. To achieve precise detection of maize pest species, this paper proposes a deep learning detection algorithm for maize pests based on an improved YOLOv8n model: Firstly, a maize pest dataset was constructed, comprising 2,756 images of maize pests, according to the types of pests and diseases. Secondly, a deformable attention mechanism (DAttention) was introduced into the backbone network to enhance the model's capability to extract features from images of maize pests. Thirdly, spatial and channel recombination convolution (SCConv) was incorporated into the feature fusion network to reduce the miss rate of small-scale pests. Lastly, the improved model was trained and tested using the newly constructed maize pest dataset. Experimental results demonstrate that the improved model achieved a detection average precision (mAP) of 94.8% at a speed of 171 frames per second (FPS), balancing accuracy and efficiency. The enhanced model can be deployed on low-power mobile devices for real-time detection, which is of great significance for the healthy development of maize agriculture.

Abstract in Chinese

玉米发生病虫害时会导致产量骤减使农民遭受重大经济损失。因此，准确、高效的玉米害虫种类检测在病虫害防治过程中可进行针对性防治。为了获得准确玉米害虫种类检测，本文提出了一种基于改进YOLOv8n的玉米害虫深度学习检测算法：首先，根据玉米病虫害种类，构建了玉米害虫数据集，共计2756张玉米害虫图像；其次，在主干网络中引入具有可变形注意力机制（DAttention）来提高算法模型对玉米害虫图像的特征提取能力；然后，针对不同尺度玉米害虫，在特征融合网络中引入空间和通道重组卷积（SCConv），降低小目标害虫的漏检率；最后，基于自建玉米害虫数据集对改进后的模型进行训练和测试。实验结果表明，改进后的模型在171帧每秒（FPS）的速度下实现了94.8％的检测平均精度（mAP），平衡了准确性和效率，可将改进的模型部署在低算力移动设备中实现实时检测，对玉米农业健康发展具有重要意义。