基于最优视图选择和VGGT架构的兰州百合种球质量、体积估测

Weight and Volume Estimation of Lanzhou Lily Bulbs Based on Next Best View and VGGT Framework

摘要:

兰州百合是我国唯一的甜百合,其种球质量、体积是影响成活率、产量及商品价值的重要因素。为解决种球形态不规则、人工表型测量费时费力的问题,本文提出了一种融合最优视图选择和VGGT架构的三维表型估测方法。利用信息熵、新点率、颜色差异、全局特征、视角冗余构建多维评价指标,筛选出由3幅图像构成的最优视图集,作为VGGT输入,端到端的生成相机参数及三维点云,并结合Solov2分割掩码进行三维投影,实现种球点云分割;进而提取14个3D形状特征,通过逐步回归消除共线性影响,构建种球质量、体积线性预测模型,而且在随机森林特征重要性筛选的基础上又构建了SVR、KNN、GBR、BPNN等非线性模型。结果表明,3个最优视图可实现90%以上的几何与颜色覆盖;BPNN质量、体积预测模型的R2均超过0.95,MAPE低于10%。另外,在GeForce RTX 3090平台,本研究提出的种球三维重建仅耗时约3 250 ms,4 s内便可完成1个样本的三维测量。研究结果为兰州百合种球自动化分选与品质评估提供了技术支撑。

Abstract:

Lanzhou lily is the only sweet lily variety in China,and the weight and volume of its bulbs are key indicators affecting survival rate,yield,and commercial value. To address the challenges of irregular bulb morphology and the low efficiency of manual phenotypic measurements,a three-dimensional phenotypic estimation method that integrated best view selection with the VGGT architecture was proposed. A multidimensional evaluation metric system was constructed based on information entropy,new-point rate,color difference,global features,and view redundancy to select an optimal set of three images as VGGT inputs. The model then performed end-to-end estimation of camera parameters and 3D point clouds. By incorporating Solov2 segmentation masks for 3D projection,precise bulb point cloud segmentation was achieved. Subsequently,totally 14 3D shape features were extracted,and stepwise regression was employed to eliminate multicollinearity effects and build linear prediction models for bulb weight and volume. Furthermore,based on feature importance ranking from a random forest model,nonlinear models,including SVR,KNN,GBR,and BPNN were developed. Experimental results demonstrated that the optimal three-view set achieved over 90% geometric and color coverage. The BPNN models for weight and volume prediction achieved R2 values above 0.95 and MAPE values below 10%. In addition,on a GeForce RTX 3090 platform,the proposed 3D reconstruction process required only took about 3 250 ms,enabling complete 3D measurement of a single sample within 4 s. The research result can provide an efficient and practical technical approach for the automated sorting and quality assessment of Lanzhou lily bulbs.