Abstract: To address the issues of low efficiency, easy false detection, missed detection and limited accuracy in pest detection based on machine vision in agricultural fields, a deep learning algorithm for pest recognition named YOLOv8—ECSI, based on an improved YOLOv8, was proposed to enhance detection performance for small, dense and similar‑featured common pests. Firstly, EfficientNetwas used to replace the original backbone network, achieving model lightweighting without significantly affecting detection accuracy. Secondly, the original upsampling module was replaced with a CARAFE upsampling module to reduce the loss of detail information. Additionally, the small object detection layer incorporated spatial‑channel reorganization convolution (SCConv) to enhance feature representation capabilities. The original loss function was replaced with the Inner—CIoU loss function to improve small object detection capabilities. A dataset was established by using the large public dataset IP102 and web queries. Experiments demonstrated that the improved model achieved an average precision of 94.6%, with a reduction of 33.6% in the number of parameters and a reduction of 23.5% in computational load. Compared with models such as YOLOv5, YOLOv7, and SSD, the improved model exhibited superior detection performance and faster speed.

Key words:  crop pests, deep learning, lightweight model, small target detection, spatial channel recombination

摘要： 针对农田中基于机器视觉的虫害检测效率低且容易出现误检、漏检、精度较低的问题，为提高对小且密集、特征接近的常见虫害的检测性能，提出一种基于改进YOLOv8的虫害识别深度学习算法YOLOv8—ECSI。首先将EfficientNet替换原有的骨干网络，在保证检测精度不受到大幅影响的前提下实现模型轻量化；其次将原始的上采样模块替换为CARAFE上采样模块，减少细节信息的损失；然后在小目标检测层加入空间通道重组卷积（SCConv），增强特征表达能力；最后将原有损失函数更换为Inner—CIoU损失函数，提高对小目标的检测能力。通过大型公开数据集IP102和网络查询建立数据集。试验表明，改进后的模型平均精度均值达到94.6%，同时参数量减少33.6%，计算量减少23.5%，比YOLOv5、YOLOv7、SSD等模型检测效果更优、速度更快。

关键词: 农作物虫害, 深度学习, 轻量化模型, 小目标检测, 空间通道重组