基于PaddlePaddle的杂草模型量化部署与结构优化研究

Quantitative deployment of weed models based on PaddlePaddle Research on Structural Optimization

针对农业中广泛应用的无人机等终端除草设备的计算资源少、存储资源有限等问题,文章通过对模型分别采取量化部署和更改模型注意力机制模块的方法来得到轻量模型。试验结果表明,移动端模型的推理时间只有服务器端模型的1/10左右,部署在树莓派3B+上的杂草分类推理速度是6fps左右,而且量化后的杂草分类模型规模有40%以上的减少,证明了量化对模型规模减少的有效性。同时基于CBAM注意力机制的MobileNet_CBAM杂草分类模型相比基于SE-Net注意力机制的MobileNetV3_Large模型在准确率上损失了0.3%,但模型参数规模降低了24.7%,整体性能更加均衡。本研究可为杂草模型的小型化落地应用提供理论基础和技术支持。

In response to the problems of low computing resources and limited storage resources in terminal weeding devices such as drones,which are widely used in agriculture,this article obtains a lightweight model by quantitatively deploying and modifying the attention mechanism module of the model.The experimental results show that the inference time of the mobile model is only about 1/10of that of the server model.The weed classification inference speed deployed on Raspberry Pi 3B+is about 6fps,and the quantified weed classification model has a reduction of over 40%in size,proving the effectiveness of quantification in reducing the model size.MobileNet based on CBAM attention mechanism simultaneously_CBAM weed classification model compared to MobileNetV3based on SE Net attention mechanism_The Large model lost 0.3%in accuracy,but the model parameter size decreased by 24.7%,resulting in a more balanced overall performance.This study can provide theoretical basis and technical support for the miniaturization and landing application of weed models.