Abstract:

Aiming at the problems of complex backgrounds， small individual size and inconsistent grading standards after harvesting of daylily buds in field environment， a recognition and post-harvest grading method for daylily buds was proposed. A database was established based on 1 716 images of daylily buds under various complex conditions such as different lighting， occlusion， and blurriness. The Biformer self-attention mechanism was introduced into the backbone network of the YOLOv5s model to train the dataset， and comparative tests were conducted against various other target detection algorithms. After recognition， a daylily bud grading algorithm was used to obtain the contours of the daylily buds through image processing technology， and geometric calculation techniques were used to measure the length and diameter of the daylily buds for grading. Experimental results showed that theimproved YOLOv5s algorithm significantly increased recognition precisionthe precision， recall rate， and mean average precision （mAP） of improved YOLOv5s algorithm of 94.80%， 91.40%， and 96.60%， respectively， in the open field recognition of daylily. The accuracy of the daylily grading algorithm reached to 97.00%， meeting the requirements for daylily grading in production practice and providing reliable support for the intelligent development of the daylily industry.

Key words: daylily recognition, biformer attention mechanism, improved YOLOv5s, open field environment, quality grading

摘要：

针对大田环境下黄花菜花蕾识别背景复杂、个体过小及采摘后分级标准不统一的问题，提出了黄花菜花蕾的识别及采摘后分级方法。选取1 716幅不同光照、遮挡及模糊等复杂环境下的黄花菜花蕾图像建立数据库，在YOLOv5s模型的主干网络中引入Biformer自注意力机制对数据集进行训练，并与多种其他目标检测算法进行对比测试。在识别完成后，使用分级算法通过图像处理技术获取黄花菜花蕾的轮廓，并使用几何计算技术获得黄花菜花蕾的长度及直径，对其进行分级。结果显示，改进的YOLOv5s算法在黄花菜的大田识别中精确率、召回率、平均精确率（mAP）分别达到了94.80%、91.40%、96.60%，识别精确率显著提高，黄花菜分级算法准确率达到97.00%，满足生产实践中对黄花菜分级的要求，为黄花菜产业智能化提供可靠支持。

关键词: 黄花菜识别, Biformer注意力机制, 改进 YOLOv5s, 大田环境, 品质分级