Pests are one of the important factors causing storage wheat losses, timely detection of pests and effective measures can reduce losses of stored wheat. The traditional artificial detection of pests method exists artificial factors have a greater impact, the problem of slow speed, based on deep learning storage grain insect detection method, although time-consuming short, but there is a larger model, speed and accuracy of the two difficult to balance the problem. Therefore, in this paper, after experiments, firstly, the YOLOv8m algorithm is selected as the basis for improvement, and then, Shufflenetv2, a lighter network, replaces Darknet—53; secondly, Squeeze—and—Excitation Networks attention mechanism is added at the end of the backbone network to obtain highquality feature maps, which effectively improves the detection accuracy. Finally, WIoUv3 Loss is adopted as the regression loss function of YOLOv8m to improve the accuracy and speed of detection. The experimental results show that the mAP of the proposed model reaches 95.4%, the number of model parameters is 19.46 M, and the FLOPs is 56.74 G. Compared with other models, the accuracy is higher, the number of model parameters is lower, and the speed is faster, which can provide effective technical support for the detection of pests in warehouses.

害虫是造成仓储小麦损失的重要因素之一，及时检测害虫并采取有效手段能够减少仓储小麦损失。传统人工检测害虫方法存在人工因素影响较大、速度慢的问题，基于深度学习的仓储粮虫检测方法虽然耗时短，但存在模型较大、速度和准确率二者难以平衡的问题。故首先选取YOLOv8m算法作为基础进行改进，接着以更轻量化的网络Shufflenetv2代替Darknet—53；其次，在主干网络末端添加Squeeze—and—Excitation Networks注意力机制获取高质量的特征图，有效提高检测精度；最后，采用WIoUv3 Loss为YOLOv8m的回归损失函数，提高检测的精度和速度。试验结果表明：所提出的改进模型平均精度均值达到95.4%，模型参数量为19.46 M，FLOPs为58.74 G。相比其他模型，精确率更高，模型参数量更低，速度更快，能够为仓储害虫检测提供有效技术支撑。