Wheat-Net: An Automatic Dense Wheat Spike Segmentation Method Based on an Optimized Hybrid Task Cascade Model

• DOI: 10.3389/fpls.2022.834938
• Publication Date: 2022-02-10T04:58:20Z
• Authors: Jiajing Zhang; An Min; Brian J. Steffenson; Wen-Hao Su; Cory D. Hirsch; James Anderson; Jian Wei; Qin Ma; Ce Yang
• Keywords: Feature (linguistics); Minimum bounding box; Bounding overwatch

Precise segmentation of wheat spikes from a complex background is necessary for obtaining image-based phenotypic information traits such as yield estimation and spike morphology. A new instance method based on Hybrid Task Cascade model was proposed to solve the detection problem with improved results. In this study, images were collected fields where environment varied both spatially temporally. Res2Net50 adopted backbone network, combined multi-scale training, deformable convolutional networks, Generic ROI Extractor rich feature learning. The methods trained validated, average precision (AP) obtained bounding box mask 0.904 0.907, respectively, accuracy counting 99.29%. Comprehensive empirical analyses revealed that our (Wheat-Net) performed well challenging field-based datasets mixed qualities, particularly those various backgrounds adjacence/occlusion. These results provide evidence dense capabilities masking, which useful not only but also morphology assessments.