文献类型：期刊文献

英文题名：Towards mechanized harvesting of pineapples: A masked self-attention instance segmentation network and pineapple detection dataset

作者：Shan, Zhe[1,2];Ye, Songtao[3];Lin, Cong[4];Xue, Zhong[1,5]

机构：[1]Chinese Acad Trop Agr Sci, South Subtrop Crop Res Inst, Zhanjiang 524088, Peoples R China;[2]Hainan Univ, Sch Comp Sci & Technol, Haikou 570228, Peoples R China;[3]Jiaotong Univ, Fac Elect & Informat Engn, Xian 710049, Peoples R China;[4]Guangdong Ocean Univ, Sch Elect & Informat Engn, Zhanjiang 524088, Peoples R China;[5]Minist Agr, Lab Trop Fruit Biol, Zhanjiang 524088, Peoples R China

年份：2025

卷号：156

外文期刊名：ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE

收录：SCI-EXPANDED(收录号：WOS:001503130400001)、、EI(收录号：20252218532417)、Scopus(收录号：2-s2.0-105006876310)、WOS

基金：This research was supported by the Hainan Province Science and Technology Special Fund, China (No. ZDYF2023XDNY058) , Top Ten Guangdong Province Agricultural Science and Technology Innovation Main Attack Directions "Unveiling and Leading" Project, China (No. 2022SDZG03) , and Central Public-interest Scientific Institution Basal Research Fund, China (No. 1630062022005) .

语种：英文

外文关键词：Smart agriculture; Deep learning; Mechanized pineapple harvesting; Grasp detection; Instance segmentation

外文摘要：Automated pineapple harvesting using robotic and computer systems has been a pressing issue that researchers need to tackle urgently. However, varying light conditions in orchards, complex environments, and leaf shading present significant challenges for accurate, real-time pineapple identification and localization. Furthermore, the damage to fruits caused by robotic harvesting further limits the development of automated pineapple harvesting. In this study, we propose a real-time instance segmentation scheme to address the above two issues simultaneously. Correspondingly, we design a masked self-attention instance segmentation network based on mixed supervised learning (MAISNet) to quickly extract pineapples' positional and geometric information and reduce the damage rate during robotic arm grasping. First, to meet the real-time needs of the robotic arm, a one-stage detection neural network is employed as the baseline model, achieving fast instance segmentation. Second, we incorporate a masked self-attention module to efficiently identify pineapple regions, reducing interference from irrelevant information. Third, we design a mixed supervised learning approach that allows the model to have some degree of uncertainty, enhancing the model's ability to recognize occluded regions while reducing the over-reliance on labels. At the same time, to promote the pineapple detection field's development and train the above algorithm, we present a public pineapple dataset. It is collected from real orchards, involves diverse complex scenarios, and is carefully labeled by hand. Many ablation experiments and comparison experiments demonstrate the validity and superiority of the method proposed in this study. Notably, experiments on edge devices verified the practical applicability of our approach in mechanized pineapple harvesting. This research changes the original paradigm of detecting only the location by accurately delineating fruit contours for direct robotic grasping. Consequently, it significantly reduces fruit damage during mechanical harvesting and advances the feasibility of large-scale automation in agriculture.