Abstract: Aiming at the deficiency of low accuracy of traditional weed identification methods for weeds with similar characteristics to forage，an improved DINO detection network for weed detection in pastures was proposed.The CBAM‑G attention mechanism module combining spatial attention and channel attention was added to the backbone network ResNet of the end‑to‑end model DINO to enhance the extraction of effective features and precise positional information，and reduce the interference of invalid information；By increasing the network depth，the backbone network could extract deeper target features；The original feature fusion module in the backbone network was replaced with the lightweight SFPN module；Finally，in order to improve the stability and detection performance of the feature extraction network and Transformer，Gaussian error linear unit was added to the model feature extraction network.The experiment showed that the AP₅₀ of the algorithm on the pasture weed data set reached 95.89%，AP₇₅reached 89.23%. Compared to the original model，it can better utilize multi‑scale feature information and improve recognition accuracy.

Key words: Weed identification, DINO, ResNet, Feature fusion, Transformer

摘要： 针对传统杂草识别方法对与牧草有相似特征的杂草的识别精度低的不足，提出一种基于改进DINO检测网络的牧场杂草检测模型。为了增强有效特征和精确位置信息的提取，并减少无效信息的干扰，在端到端模型DINO的主干网络ResNet中加入结合空间注意力和通道注意力的CBAM-G注意力机制模块；通过增加网络深度，让主干网络可以提取到更深层次的目标特征；引入更加轻量化的SFPN模块，替换了算法中原有的特征融合模块；最后为了提高特征提取网络与Transformer的稳定度和检测性能，在模型特征提取网络中加入高斯误差线性单元。结果表明，改进后的检测模型在Kaggle的牧场杂草数据集上的像素精度AP₅₀达到了95.89%，AP₇₅达到了89.23%，相较于原始模型可以更好地利用多尺度特征信息，并提升识别精度。

关键词: 杂草识别, DINO, ResNet, 特征融合, Transformer