基于空间和通道注意力的花椒图像枝条分割  

SPATIAL AND CHANNEL ATTENTION-BASED BRANCH SEGMENTATION FOR PEPPER IMAGE

作　　者：刘兴科 杨前 罗建桥[1] 刘雪垠[2] 李柏林[1] Liu Xingke;Yang Qian;Luo Jianqiao;Liu Xueyin;Li Bailin(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,Sichuan,China;R&D Center,Sichuan Mechanical Research and Design Institute,Chengdu 610063,Sichuan,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031 [2]四川省机械研究设计院研发中心,四川成都610063

出　　处：《计算机应用与软件》2025年第3期202-208,243,共8页Computer Applications and Software

基　　金：四川省科技计划重点研发项目(2021YFN0020)。

摘　　要：提取主干枝条是花椒视觉采摘的关键问题。针对现有网络模型难以从复杂自然场景中分割出近景主干枝条的问题,向网络中加入新设计的注意力模块,提出基于空间和通道注意力的RGBD图像花椒枝条分割模型。首先,由两个编码器分别提取彩色和深度图像特征。然后,设计空间注意力模块融合这两种特征。空间权值同时考虑外观和距离信息,能够抑制远景枝条的干扰。最后,构造通道注意力模块关注枝条主要形状信息,避免枝条细节轮廓的干扰。实验表明,相比现有双编码器网络,所提模型的分割准确率、交并比分别提高了14.72%、17.65%。所提模型能够分割出满意的近景主干枝条。Extracting the main branches is the key issue for visual picking of pepper.Aimed at the problem that the existing network model is difficult to segment the close-range main branches from complex natural scenes,a newly-designed attention module is added to the network,and a RGBD image pepper branch segmentation model based on spatial and channel attention is proposed.The color and depth image features were extracted by two encoders respectively.The design space attention module combined these two features.The spatial weight considered appearance and distance information at the same time,which could suppress the interference of distant branches.The channel attention module was constructed to pay attention to the main shape information of the branches and avoid the interference of the details of the branches.Experiments show that compared with the existing dual-encoder network,the segmentation accuracy and cross-union ratio of the proposed model are increased by 14.72%and 17.65%,respectively.The proposed model can segment satisfactory close-range trunk branches.

关 键 词：花椒视觉采摘RGBD图像分割 深度学习 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]