CAJ | 학술논문

基于图像序列的植株三维结构重建是植物无损测量的重要方法之一。而对重建模型的精度评估方法大多基于视觉逼真程度和常规测量数据。该研究以精确的激光扫描三维模型为参照，采用豪斯多夫距离，从三维尺度上对基于图像序列的植株三维重建模型进行精度评估。同时，从植株表型参数（叶片长、宽、叶面积）方面，对植株三维重建模型进行精度评估。结果表明，基于图像序列的三维重建模型精度较高，豪斯多夫距离在0～10 mm之间，各试验植株豪斯多夫距离大多小于4.0 mm，各植株表型参数与其对照值的R2均大于0.95，且两者的无显著性差异（P＜0.05）。此植株三维结构重建方法能够应用于植物表型、基因育种、植物表型与环境互作等研究领域。
기우도상서렬적식주삼유결구중건시식물무손측량적중요방법지일。이대중건모형적정도평고방법대다기우시각핍진정도화상규측량수거。해연구이정학적격광소묘삼유모형위삼조，채용호사다부거리，종삼유척도상대기우도상서렬적식주삼유중건모형진행정도평고。동시，종식주표형삼수（협편장、관、협면적）방면，대식주삼유중건모형진행정도평고。결과표명，기우도상서렬적삼유중건모형정도교고，호사다부거리재0～10 mm지간，각시험식주호사다부거리대다소우4.0 mm，각식주표형삼수여기대조치적R2균대우0.95，차량자적무현저성차이（P＜0.05）。차식주삼유결구중건방법능구응용우식물표형、기인육충、식물표형여배경호작등연구영역。
Plant architecture is an important determinant of the canopy light interception and photosynthesis. Therefore, effective and nondestructive methods for obtaining plant architecture can help us understand the relationships between plant physiological processes and morphogenesis. Digital camera technology has become relatively ubiquitous and inexpensive, leading to a recent surge in utilizing plant imaging to capture data. Therefore, a three-dimensional (3-D) reconstruction of plant architecture based on these photographed image sequences can be realized. However, the accuracyevaluation of reconstruction is always determined from visual effect and or from 1-D or 2-D measured data. In this study, image sequences were obtained around experimental plants (e.g. egg plant, sweet pepper and cucumber) by slightly moving a commercial camera for image generation so that each neighboring image pair shared short baseline. Structure from motion (SFM) method was applied to produce a set of sparse point cloud based on plant image sequences. As the sparse point cloud was inadequate for the reconstruction of complicated plant architecture, multiple-view stereo (MVS) method was further used to produce dense and accurate point cloud based on the output of SFM. Software Bundler and CMVS were applied to implement the SFM and MVS methods, respectively. Bundlertakes a set of images as input, and produces a 3D reconstruction of sparse scene geometry and camera parameters as output. CMVS takes a set of images and camera parameters (the output of Bundler) as input, and outputs a set of dense points with geometry details. Once original point cloud has been obtained, point cloud processing procedures were conducted to refine point cloud, including deleting noise points and scaling to the actual size of experimental plants. In order to obtain point cloud of individual blade, segmentation of point cloud of plants was conducted based on region growing segmentation algorithm. Once point cloud of individual blade was obtained, Poisson surface reconstruction algorithm was used to reconstruct each leaf. Before accuracy evaluation, point cloud of individual leaf blade and corresponding laser scanning point cloud were aligned to the same 3-D coordinate system using the Iterative Closest Point algorithm. Then, comparison on 3-D scale based on Hausdorff distance was made between point cloud data obtained from plant image sequences and referenced point cloud data with laser scanning on individual leaf blade level. Furthermore, phenotypic attributes, such as leaf blade width, leaf blade length and blade area were extracted based on data from image sequences and from laser scanning methods. Besides, these attributes of each blade were manually measured. Finally, comparisons were made for blade area, blade length and blade maximum width between data from the image sequence-based model, laser scanning based model and manually measured data. The results showed that high accuracy of 3-D reconstruction was obtained based on plant image sequence method. Hausdorff distances of experimental plants were ranged from 0 to 10 mm, and most of the values were less than 4.0 mm. There was a good agreement between measured and calculated blade area, blade length and maximum width withR2> 0.95 for blade area, RMSE< 4.5 mm for blade width, and RMSE < 5.6 mm for blade length. There was no significant difference for each attribute between measured and calculated data (ANOVA,P> 0.05). A key advance of the current 3-D reconstruction of plant architecture is the capability to non-destructively capture plant traits with high accuracy. This advance permits time-series measurements that are necessary to follow the progression of growth and stress on individual plants, and will play an important role in related research fields, such as plant phenotyping, genetic breeding, interactions between plant phenotype and environment.