基于TD-Mask R-CNN的机械装配体图像实例分割
Instance Segmentation of Mechanical Assembly Image Based on TD-Mask R-CNN
唐若仪 1陈成军 1王金磊 1代成刚1
作者信息
- 1. 青岛理工大学机械与汽车工程学院,青岛 266520
- 折叠
摘要
在机械产品装配过程中,为了准确识别机械装配体零件信息以减少零件漏装、错装等现象,提出一种改进的机械装配体图像实例分割方法TD-Mask R-CNN.首先,在主干网络ResNet101 中引入可变形卷积(deformable convolutional networks,DCN)以增加网络模型的泛化能力;其次,使用Transfiner结构作为掩码分支以提高机械零件边缘的分割精度;最后,在Transfiner结构中引入离散余弦变换(discrete cosine transform,DCT)模块以提升模型对机械装配体图像整体的分割能力.实验结果表明,提出的实例分割方法在合成深度图像数据集和真实彩色图像数据集上得到的掩码平均精度(average precision,AP)分别为87.7%和92.0%,与其他主流实例分割算法相比均有所提升.
Abstract
In the assembly process of mechanical products,in order to accurately identify the mechanical parts information to reduce the phenomenon of parts leakage and misassembly,this paper proposes an im-proved instance segmentation method TD-Mask R-CNN for mechanical assemblies images.First,deformable convolutional networks(DCN)is introduced into the backbone network ResNet101 to increase the general-ization ability of the network model.Then,the Transfiner structure is used as a mask branch to improve the segmentation accuracy of the mechanical part edges.Finally,the discrete cosine transform(DCT)module is introduced in the Transfiner structure to improve the model's ability to segment the whole image of the mechanical assembly.The experimental results show that the average precision(AP)of the mask obtained by this method on the synthetic depth image dataset and the real color image dataset are 87.7％ and 92. 0％ ,respectively,which is improved compared with other mainstream instance segmentation algorithms.
关键词
深度学习/装配监测/实例分割/TD-Mask/R-CNNKey words
deep learning/assembly monitoring/instance segmentation/TD-Mask R-CNN引用本文复制引用
基金项目
国家自然科学基金项目(52175471)
山东省自然科学基金项目(ZR2021MF110)
出版年
2024
NETL
NSTL
维普
万方数据