基于多维表型特征的盆栽白雪凤梨分级研究

Grading Research for Potted Snow Pineapple Based on Multidimensional Phenotypic Traits

摘要:

针对盆栽花卉白雪凤梨人工分级效率低、一致性差的问题,本文提出一种基于轻量化分割模型与多维表型特征提取的智能分级方法,并研制配套的自动化分级装备。为提升凤梨植株分割的实时性与准确性,对YOLO v8n-seg 模型进行结构改进,引入轻量化REPCSP(Reparameterized cross stage partial)模块替代C2f(CSP bottleneck with 2 convolutions)模块,并采用SCSHead(Shared convolutional segmentation head)替代传统分割检测头,构建YOLO v8n-RS 模型,改进后模型在分割平均精度均值(mAP)达98. 0%的同时,帧率提升至1 121. 8 f/ s,参数量仅为1. 7 ×106 ,帧率较YOLO v8n-seg 提升30. 8%,更适用于实时分级场景;为实现白雪凤梨的智能分级需求,结合模型分割结果与深度图像处理技术,实现株高、花径、倾斜角与花朵饱满度等关键表型参数的精准提取,结果表明:株高、花径、倾斜角平均绝对误差分别为0. 745 cm、0. 410 cm 和1. 627°,对应R2 分别为0. 933、0. 915 和0. 773,花朵饱满度等级分类准确率为85. 3%,F1 值为84. 9%;基于所提方法与系统,研制自动化分级装置并开展100 株白雪凤梨实地分级试验,系统分级准确率达89. 0%,单株平均处理时间为3. 54 s,验证了智能分级系统在实际生产环境中的高效性、稳定性。

Abstract:

Aiming to address the issues of low efficiency and poor consistency in manual grading of potted ornamental plants such as Ananas comosus var. bracteatus (‘White Snow Pineapple’), an intelligent grading method was proposed based on a lightweight segmentation model and multi-dimensional phenotypic feature extraction, and a corresponding automated grading equipment was developed. To enhance the real-time performance and accuracy of plant segmentation, the YOLO v8n-seg model was structurally optimized by introducing the lightweight reparameterized cross stage partial ( REPCSP) module to replace the original CSP bottleneck with 2 convolutions (C2f) module and adopting shared convolutional segmentation head (SCSHead) in place of the conventional segmentation head, resulting in the YOLO v8n-RS model. The improved model achieved a segmentation accuracy (mAP) of 98. 0%, with an inference speed of 1 121. 8 f/ s, and only 1. 7 ×106 parameters. Compared with YOLO v8n-seg, the inference speed was increased by 30. 8%, making it more suitable for real-time grading scenarios. To meet the requirements of intelligent grading for Snow Pineapple, key phenotypic traits, including plant height, flower diameter, inclination angle, and flower fullness were accurately extracted by combining the segmentation results with depth image processing techniques. Experimental results showed that the mean absolute errors for plant height, flower diameter and inclination angle were 0. 745 cm, 0. 410 cm and 1. 627°, respectively, with corresponding R2 values of 0. 933, 0. 915 and 0. 773. The flower fullness classification achieved an accuracy of 85. 3% and an F1-score of 84. 9%. Based on the proposed method and system, an automated grading device was developed, and a field test involving 100 Snow Pineapple plants was conducted. The system achieved a grading accuracy of 89. 0%, with an average processing time of 3. 54 s per plant, demonstrating its high efficiency and stability in real-world production environments.