基于YOLO–Fusion算法的甘蓝移栽机取苗系统设计与试验

Design and Testing of Seedling Extraction System for Kale Transplanters Based on YOLO – Fusion Algorithm

摘要:

针对甘蓝移栽机取苗时育苗盘钵苗相互遮挡导致的取苗困难,及投苗时残次苗引发的漏栽等问题,结合甘蓝钵苗尺寸与取投苗流程,设计了基于深度学习的移栽机智能取苗系统。该系统采用YOLO–Fusion多模型协同算法,由YOLO v5s–CCB位置检测算法和YOLO v8s–cls残次苗分类算法构成,分别实现茎秆位置检测及残次苗筛选。其中,YOLO v5s–CCB是在YOLO v5s模型基础上改进而来,通过在主干网络的C3模块中融入坐标注意力机制,增强对茎秆特征的提取能力;并将颈部网络中的Concat操作替换为加权双向特征金字塔网络(Bidirectional feature pyramid network,BiFPN),提升特征自适应融合能力,减少因茎叶特征相似度导致的误判。对比试验表明,YOLO v5s–CCB相较于基准模型,虽召回率和帧率分别降低0.7个百分点和11.9 f/s,但主要目标P0的精确率与整体平均精度均值分别提升2.4、2.8个百分点,达91.7%、60.5%;与RT–DETR、Faster R–CNN和YOLO v8s等主流模型相比,YOLO v5s–CCB保持更高检测精度与更快检测速度。YOLO v8s–cls算法通过识别残次苗特征,做出壮苗保留与残次苗剔除的决策,经测试集验证,理论分类成功率达到85.1%。在搭建的试验平台上部署YOLO–Fusion算法,开展茎秆位置映射与取投苗试验:位置映射试验中,z方向上的实际坐标与映射坐标最大误差为5.1 mm,满足精准取苗的需求;取苗试验中,系统采用YOLO–Fusion算法后,平均取苗成功率达83.3%,较未采用时提升26.4个百分点;分类投苗试验中,对壮苗和残次苗判断准确率分别达到94.5%和70.8%,平均分类成功率为82.6%。因此,YOLO–Fusion算法可有效提高取苗成功率与残次苗分类效率,为智能甘蓝移栽机后续研究提供理论与技术支撑。

Abstract:

Aiming at the problems such as the difficulty of picking up seedlings due to the mutual obstruction of seedlings in the nursery tray when picking up seedlings by kale transplanter and the omission of seedlings caused by defective seedlings during seedling casting, the intelligent seedling picking system of transplanter based on deep learning was designed by combining the size of kale potting seedling with the process of picking up and casting seedlings. The system adopted the YOLO – Fusion multi-model collaborative algorithm, which consisted of the YOLO v5s – CCB position detection algorithm and the YOLO v8s – cls residual seedling classification algorithm to achieve the tasks of stalk position detection and residual seedling screening, respectively. Among them, YOLO v5s – CCB was improved on the basis of YOLO v5s model, which enhanced the extraction ability of stem features by incorporating the coordinate attention mechanism in the C3 module of the backbone network; and replaced the Concat operation in the neck network with bidirectional feature pyramid network (BiFPN) to improve the feature adaptive fusion ability and reduce the misjudgement due to the similarity of stem and leaf features. Comparison tests showed that YOLO v5s – CCB, compared with the benchmark model YOLO v5s, although Recall (R) and frame rate ( FPS) were reduced by 0.7 percentage point and 11.9 f/s, the precision of the main target P0 and the overall mean average precision (mAP) were improved by 2.4 and 2.8 percentage points to 91.7% and 60.5%, respectively; and compared with the RT – DETR, Faster R – CNN and YOLO v8s, YOLO v5s – CCB maintained higher detection accuracy and faster detection speed compared with mainstream models. YOLO v8s – cls algorithm made the decision of strong seedling retention and residual seedling rejection by identifying residual seedling features, which was verified in the test set, and the theoretical classification success rate reached 85.1%. The YOLO – Fusion algorithm was deployed on the experimental platform to carry out stem position mapping and seedling picking and dropping tests: in the position mapping test, the maximum error between the actual coordinates and the mapped coordinates in the z direction was 5.1 mm, which met the demand for accurate seedling picking; in the seedling picking test, the average seedling picking success rate of the system using the YOLO – Fusion algorithm reached 83.3%, which was 26.4 percentage points higher than that of the system when the algorithm was not used; the success rate of seedling sorting and dropping reached 85.1% after verification by the test set. In the seedling sorting test, the accuracy of judging strong and defective seedlings reached 94.5% and 70.8%, respectively, and the average sorting success rate was 82.6%. Therefore, the YOLO – Fusion algorithm could effectively improve the success rate of seedling picking and the classification efficiency of defective seedlings, and provide theoretical and technical support for the subsequent research of intelligent kale transplanting machine.