Fig 1.
Overview of canola silique segmentation and counting.
(A) Sparse reconstruction. (B) Dense reconstruction. (C) Point cloud preprocessing. (D) Point cloud augmentation. (E) Segmentation model prediction result. (F) Segmented Silique Point Cloud. (G) Silique instance clustering result.
Fig 2.
Imaging platform.
Table 1.
Rapeseed video acquisition and image processing data protocol.
Fig 3.
Reconstructed canola point cloud samples.
Fig 4.
KAN-GLNet network model.
Fig 5.
The overall architecture of KGL-PointNet and its core components.
(A) The complete network structure of KGL-PointNet. (B) The multi-scale feature enhancement module, GLS Attention. (C) The feedforward network based on partial convolution, PCFN.
Fig 6.
Network architecture diagrams of Kolmogorov-Arnold, FastKAN, and reverse bottleneck KAN convolutions.
(A) Kolmogorov-Arnold and FastKAN Convolutions. (B) Reverse Bottleneck KAN Convolutions.
Fig 7.
Network architecture diagrams of the GS attention module and the LS attention module.
(A) GS Attention module structure. (B) LS Attention module structure.
Table 2.
Definitions of evaluation metrics used for semantic segmentation and clustering.
Table 3.
Hardware and software specifications.
Fig 8.
Training loss and accuracy curves of KAN-GLNet.
Table 4.
A comparison of semantic segmentation performance across five networks, with the top results highlighted in bold.
Fig 9.
KAN-GLNet and baseline models segmentation visualization on test set.
Table 5.
Comparison between the actual number and the detected number of canola siliques.
Table 6.
Ablation experiments on KAN-GLNet were conducted using different modules and combinations on the testing set.
Table 7.
Comparison of segmentation accuracy and parameter count between Kolmogorov-Arnold convolutions, FastKAN convolutions, reverse bottleneck KAN convolutions, and the baseline model.
Table 8.
Comparison of reverse bottleneck KAN convolutions replacing single-layer and multi-layer convolutions (Conv).
Fig 10.
(A) Impact of varying parameters on clustering performance (ground truth silique count: 78). (B) Impact of varying
parameters on clustering performance (ground truth silique count: 144).
Fig 11.
Comparison of DBSCAN clustering with and without SOR filtering.
(A) Original silique labels. (B) Clustering result using DBSCAN. (C) Clustering result using DBSCAN with SOR filtering.
Fig 12.
Canola experimental field at Sichuan Agricultural University, Ya’an City, Sichuan Province, China.