Abstract: Field weed eradication is the premise of promoting stable agricultural production. The diversity of weed species and the diversity of size, color, and location of the same species result in the unsatisfactory performance of traditional weed detection algorithms. A field weed detection algorithm based on multiscale attention and depth separable convolution is proposed in this paper. Firstly, the backbone network VGG-16 is improved by deep separable convolution to reduce the number of model parameters and speed up the model training. Then, the multiscale attention module is used to extract the multiscale features of weeds to enhance the ability of the proposed model to capture morphological image features. The results show that the precision of the proposed algorithm is 9469%, the recall is 9488%, and the F1 is 9382%. Compared with the current mainstream weed detection models, the proposed algorithm has a lower time overhead while maintaining higher detection performance. It can be used for the automatic detection of weeds in farmland.

Key words: field weed recognition, depth separable convolution, multiscale attention, morphological image features

摘要： 农田杂草根除是促进农业稳定生产的前提。由于杂草种类多，且相同物种因大小、颜色和位置的变化多样，导致传统农田杂草检测算法性能不高。提出一种基于多尺度注意力和深度可分离卷积的农田杂草检测算法。首先，利用深度可分离卷积改进主干网络VGG-16，降低模型参数量，加快模型的训练；然后，采用多尺度注意力模块提取杂草的多尺度特征，增强模型对形态图像特征的捕获能力。通过在不同时间段测试多个农田杂草样本，结果表明：本文算法的精准率为9469%、召回率为94.88%和F1值为93.82%。与当前主流杂草检测模型相比，在保持较高检测性能的基础上，具有更低的时间开销，可应用于农田杂草的自动检测。

关键词: 农田杂草检测, 深度可分离卷积, 多尺度注意力, 形态图像特征