Abstract: The text on the packaging of agricultural materials contains the registration certificate number, active ingredient content, production license number, product standard number and other producerelated information, which not only provides an important basis for farmers to buy agricultural materials, but also helps agricultural materials supervision agencies to find problems affecting the safety of agricultural materials, as well as greatly helps the identification of agricultural materials for export. Based on the data set of agricultural materials packaging images, a text detection model of agricultural materials labels based on attention mechanism was proposed. The model uses Swin-Transformer as the backbone network, uses FPN to extract text features, designs Twin Feature Fusion Module (TFFM) to integrate local features and global features, and uses scaling expansion algorithm to generate text borders in the prediction stage. The experimental results of the model on the selfbuilt agricultural materials packaging image data set show that the accuracy, recall rate and F value of the model are 91.4%, 87.3% and 89.3%, respectively, which are superior to the mainstream methods and have certain advantages for agricultural materials packaging image text detection task.

Key words: computer vision, agricultural materials labels, semantic segmentation, text detection, text recognition

摘要： 农资包装上的文本含有登记证号、有效成分含量、生产许可证号、产品标准号等产品相关信息，这些不仅为农民购买农资提供重要的依据，还有助于农资监督机构发现影响农资安全的问题，同时也对出口农资的识别有极大的帮助。基于农资包装图像构建数据集，提出一种基于注意力机制的农资标签文本检测模型，该模型使用Swin-Transformer作为骨干网络，采用FPN提取文本特征，设计双特征融合模块（Twin Feature Fusion Module，TFFM）来统合局部特征和全局特征，预测阶段采用缩放式扩展算法来生成文本边框。该模型在自建农资包装图像数据集上的试验结果表明：其准确率、召回率和F值分别为91.4%、87.3%和89.3%，均优于主流方法，对农资包装图像文本检测任务具有一定的优越性。

关键词: 计算机视觉, 农资标签, 语义分割, 文本检测, 文本识别