基于R2UNet和空洞卷积的羊后腿分割目标肌肉区识别  

Target Muscle Region Recognition in Ovine Hind Leg Segmentation Based on R2UNet and Atrous Convolution Algorithm

作　　者：刘楷东 谢斌[1,2] 翟志强 温昌凯[1,2] 侯松涛 李君 LIU Kaidong;XIE Bin;ZHAI Zhiqiang;WEN Changkai;HOU Songtao;LI Jun(College of Engineering,China Agricultural University,Beijing 100083,China;Beijing Key Laboratory of Optimized Design for Modern Agricultural Equipment,China Agricultural University,Beijing 100083,China)

机构地区：[1]中国农业大学工学院,北京100083 [2]中国农业大学现代农业装备优化设计北京市重点实验室,北京100083

出　　处：《农业机械学报》2020年第S02期507-514,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2018YFD0700804)。

摘　　要：针对前处理工序造成的羊肉智能精细分割目标肌肉区图像识别准确度低的问题,以羊后腿自动去骨分割工序为研究对象,提出一种基于R2UNet和紧凑空洞卷积的羊后腿分割目标肌肉区识别方法。对传统的UNet语义分割网络进行改进,以UNet为骨架网络,采用残差循环卷积块替换原始UNet的特征编码模块和解码模块中的卷积块以避免UNet的梯度消失,在特征编码模块和特征解码模块之间增加一个紧凑的四分支空洞卷积模块对语义特征进行多尺度编码,实现缝匠肌图像分割模型的构建。一方面,针对缝匠肌这一核心目标肌肉区,采集羊后腿图像构建数据集训练与测试本文模型,以验证该方法的准确性与实时性;另一方面,通过旋量法标定夹爪坐标系、相机点云坐标系、机器人坐标系的齐次变换矩阵以计算分割路径,并采用主动柔顺的力/位混合控制方法操纵分割机器人进行目标切削运动,验证基于本文方法得到的目标图像开展目标肌肉分割的可行性。相关试验结果表明:当交并比为0.8588时,本文方法平均精确度为0.9820,优于R2UNet的(0.8324,0.9775);单样本检测时间平均为82 ms,说明本文方法可快速、准确分割出缝匠肌图像,满足机器人自主分割系统的实时性要求,优于UNet、R2UNet、AttUNet算法。最后,在本文方法得到的缝匠肌图像基础上开展机器人实机分割试验,机器人对5条羊后腿的平均切削时间为7.9 s,平均偏移距离为4.36 mm,最大偏移距离不大于5.9 mm,满足羊后腿去骨分割的精度要求。Research on ovine meats intelligent segmentation remains limited because of the low recognition accuracy of target muscle region image caused by pre-processing process.A method of target muscle region recognition in ovine hind leg intelligent segmentation based on R2UNet and dense atrous convolution algorithm was presented.The traditional UNet semantic segmentation network was taken as the backbone network and improved.The convolution blocks in the feature encoder and decoder of the original UNet were replaced with the residual recurrent convolution blocks to avoid the gradient loss of the UNet and a four branch dense atrous convolutional module was added between the feature encoder and the feature decoder to code multi-scale semantic features.On the one hand,aiming at the sartorius muscle region,the ovine hind leg images were collected to build a dataset and the model was trained and tested using the dataset to validate the accuracy and real-time performance of this method;on the other hand,the homogeneous transformation matrix of gripper coordinate system,camera point clouds coordinate system and robot coordinate system was calibrated based on screw theory to calculate the segmentation path,and the robot cutting manipulation was controlled by an active compliant force/position hybrid control method,which validated the feasibility of target muscle segmentation based on the target image obtained by this method.The experimental results showed that when the intersection over union(IOU)was 0.8588,the average precision(AP)of the proposed method was 0.9820,which was better than that of R2UNet(0.8324,0.9775);the average time of single sample detection was 82 ms,which showed that this method can segment the sartorius image quickly and accurately,which met the real-time requirements of robot autonomous segmentation system,and it was better than UNet,R2UNet and AttUNet algorithms.Finally,based on the image of sartorius muscle obtained by this method,the real robot segmentation experiment was carried out.The average time o

关 键 词：羊后腿 自主分割 目标肌肉区识别 语义分割 空洞卷积 残差神经网络 

分 类 号：TP251.5+3[自动化与计算机技术—检测技术与自动化装置] TP183[自动化与计算机技术—控制科学与工程]