全文下载:
202603018.pdf
文章编号:1672-6987(2026)03-0148-11;DOI:10. 16351/j. 1672-6987. 2026. 03. 018
田殿雄 1 ,赵文炎 1* ,钟 诚 1 ,张 欢 1 ,李 勇 2 (1. 国网冀北电力有限公司唐山供电公司,河北 唐山 263000; 2. 上海洲固电力科技有限公司,上海 200000)
摘 要:在物流仓库自动化管理领域,机器人技术的快速发展对目标检测、语义分割和抓取点 检测的协同性能提出了更高要求。本文提出了一种多任务动态感知协同反卷积神经网络 (MTDPS-DNN),通 过 动 态 任 务 感 知 反 卷 积 模 块(DTD)和 跨 任 务 协 同 注 意 力 融 合 模 块 (CSAF),将目标检测、语义分割和抓取检测三大任务高效集成于单一模型中。DTD 模块通 过自适应特征融合策略优化了多尺度细节提取,而 CSAF 模块则增强了任务间的语义协同,显 著提升了模型的整体性能。在亚马逊机器人挑战赛等权威数据集上的实验表明,该模型在目 标检测中的全局准确率达到 92. 26%,平均 IoU 为 83. 08%,语义分割的 mIoU 为 76. 01%,抓取 点平均距离仅为 26. 84 像素,性能优于主流模型如 YOLOv5、DETR 和 DeepLabv3+。消融实 验进一步验证了多任务学习策略的有效性,尤其在复杂场景中表现出更强的鲁棒性。抓取实 验证实了该模型的实用性和稳定性,为物流仓储和电力设备维护提供了可靠的技术支持。
关键词:目标检测;语义分割;抓取点检测;多任务
中图分类号:TP 391 文献标志码:A
引用格式:田殿雄,赵文炎,钟诚,等 . 面向多任务的动态感知协同反卷积网络[J]. 青岛科 技大学学报(自然科学版),2026,47(3):148-158.
TIAN Dianxiong, ZHAO Wenyan, ZHONG Cheng, et al. Multi-task dynamic perception syn⁃ ergistic deconvolution network[J]. Journal of Qingdao University of Science and Technology (Natural Science Edition),2026,47(3):148-158.
Multi-Task Dynamic Perception Synergistic Deconvolution Network
TIAN Dianxiong1 ,ZHAO Wenyan1 ,ZHONG Cheng1 ,ZHANG Huan1 ,LI Yong2 (1. Tangshan Power Supply Company of State Grid Jibei Electric Power Co. , Tangshan 263000, China; 2. Shanghai Zhougu Power Technology Co. , Ltd. , Shanghai 200000, China)
Abstract:In the field of logistics warehouse automation management, the rapid development of robotics has put forward higher requirements on the collaborative performance of target detec⁃ tion, semantic segmentation and grasping point detection. In this paper, a multi-task dynamic perceptual synergistic deconvolutional neural network (MTDPS-DNN) is proposed to effi⁃ ciently integrate three major tasks, namely, target detection, semantic segmentation, and grasping detection, into a single model by means of a dynamic task-aware deconvolution module (DTD) and a cross-task synergistic attention fusion module (CSAF). The DTD module optimizes multiscale detail extraction by means of an adaptive feature fusion strategy while the CSAF module enhances the semantic synergy between tasks, significantly improving the overall performance of the model. Experiments on authoritative datasets such as Amazon Robotics Challenge show that the model achieves a global accuracy of 92. 26% in target detec⁃ tion, an average IoU of 83. 08%, an mIoU of 76. 01% for semantic segmentation, and an aver⁃ age distance of only 26. 84 pixels from the grasping point, which outperforms mainstream models such as YOLOv5, DETR, and DeepLabv3+ . The ablation experiments further vali⁃ date the effectiveness of the multi-task learning strategy, especially in complex scenes that show stronger robustness. The grasping experiment confirms the practicality and stability of the model, which provides reliable technical support for logistics warehousing and power equip⁃ ment maintenance.
Key words:target detection; semantic segmentation; grasping point detection; multi-task
收稿日期:2025-07-20
基金项目:山东省自然科学基金项目(ZR2024MF142);国网冀北电力有限公司唐山供电公司基金项目(SGJBTS00WZJS2311451).
作者简介:田殿雄(1988—),男,工程师 . * 通信联系人 .