计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (18): 263-269.DOI: 10.3778/j.issn.1002-8331.2005-0438

• 工程与应用 • 上一篇    下一篇

改进DeepLabv3+和XGBoost的羊骨架切割方法

李振强,王树才,赵世达,白宇   

  1. 华中农业大学 工学院,武汉 430070
  • 出版日期:2021-09-15 发布日期:2021-09-13

Cutting Methods of Sheep’s Trunk Based on Improved DeepLabv3+ and XGBoost

LI Zhenqiang, WANG Shucai, ZHAO Shida, BAI Yu   

  1. College of Engineering, Huazhong Agricultural University, Wuhan 430070, China
  • Online:2021-09-15 Published:2021-09-13

摘要:

为实现羊骨架自动切割,提出一种基于DeepLabv3+和XGBoost的羊骨架切割方法。该方法通过研究DeepLabV3+网络架构,基于ResNet-101搭建了4种基础网络,通过调整空洞卷积的扩张率和引入可形变卷积核的方法设计改进了2种ASPP结构。共搭建8种羊骨架特征部位分割网络,按照6∶2∶2的比例划分数据集。与DeepLabv3+进行对比实验,优化后DeepLabv3+的mIoU、PA和F值分别为0.849、0.870和0.879,能够较好地实现羊骨架特征部位分割。基于分割结果对羊骨架特征部位进行特征提取,共获得35组形位特征参数,对特征集进行归一化等预处理操作。基于XGBoost搭建羊骨架切割位置预测模型,模型均方根误差MSE为8.18,拟合度R2为0.949,坐标残差绝对平均值为2.47像素点,模型具有较强的预测能力和泛化能力。基于机器人平台进行切割实验,采用3组样本进行对比实验,羊骨架切割精度为3.25?mm,理论效率为413只/h,约提升37.9%,结果表明该方法有效可行且具备较高精度。

关键词: 羊肉架切割, DeepLabv3+, 图像分割, 特征提取, XGBoost

Abstract:

To realize the automatic cutting of sheep’s trunk, the cutting methods of sheep’s trunk based on DeepLabv3+ and XGBoost is proposed. By studying DeepLabv3+ network architecture, four basic networks are designed based on ResNet-101. The ASPP structure is improved by modifying expansion rate and introducing deformable convolution. Eight kinds sheep’s trunk segmentation networks are built, and the data sets are divided according to the ratio of 6∶2∶2, to compare with DeepLabv3+. The mIoU, PA and F value of the optimized DeepLabv3+ are 0.849, 0.870, 0.879, which can better meet the requirements of the segmentation task of the characteristic parts of the sheep’s trunk. Based on the image segmentation, the features of the skeleton are extracted. A total of 35 configuration bit feature parameters are obtained, and the feature set is normalized and processed. Prediction model is established based on XGBoost, the MSE and R2 are 8.18 and 0.949, the mean absolute value of coordinate residual is 2.47. It shows that the model has strong prediction ability and generalization ability. The experiment is carried out on robot platform. The error is about 3.25 mm and the theoretical efficiency has reached 413 units per hour, which increased by 37.9%. The results show that the method is feasible and has high precision.

Key words: the cutting of sheep’s trunk, DeepLabv3+, image segmentation, feature extraction, XGBoost