欢迎访问《图学学报》 分享到:

图学学报 ›› 2021, Vol. 42 ›› Issue (4): 572-580.DOI: 10.11996/JG.j.2095-302X.2021040572

• 图像处理与计算机视觉 • 上一篇    下一篇

基于深度学习的主轴承盖分类识别算法

  

  1. 1. 武汉理工大学机电工程学院,湖北 武汉 430070;
    2. 广州工业大学机电工程学院,广东 广州 510006
  • 出版日期:2021-08-31 发布日期:2021-08-05

Classification algorithm of main bearing cap based on deep learning

  1. 1. School of Mechanical and Electrical Engineering, Wuhan University of Technology, Wuhan Hubei 430070, China;
    2. School of Mechanical and Electrical Engineering, Guangdong University of Technology, Guangzhou Guangdong 510006, China
  • Online:2021-08-31 Published:2021-08-05

摘要: 机械零件自动分类识别算法,在智能工业、自动化加工等领域具有广阔地应用前景。针对汽车
发动机主轴承盖零件自动分类时,存在特征多表面分布和光照敏感等难点问题,提出多分支特征融合卷积神经
网络(MFF-CNN)。MFF-CNN 具有 2 个子网络分支,分别提取主轴承盖 2 个表面的特征,经过特征融合,形成
最终的零件分类特征。在网络结构设计上,MFF-CNN 基于密集连接型卷积神经网络设计,通过增强网络层级
间的特征重用,有效降低模型的参数量,缓解较小样本量条件下,深层网络的过拟合和梯度消失问题。实验结
果表明,在实际采集的主轴承盖图像数据集上,MFF-CNN 的识别率为 91.6%,并对实际生产中的零件图像光
照不均匀问题,具有良好的鲁棒性。

关键词: 机械零件识别, 卷积神经网络, 细粒度图像分类, 特征融合

Abstract: The automatic classification and recognition algorithm of mechanical parts has broad application prospects
in the fields of intelligent industry and automatic processing. In the automatic classification of automobile engine
main bearing cap parts, there are difficult problems such as multi-surface distribution of features and light sensitivity.
A multi-branch feature fusion convolutional neural network (MFF-CNN) was designed. The two sub-network
branches of the MFF-CNN can extract the features of the two surfaces of the main bearing cap respectively, and form
the final part classification feature after feature fusion. In terms of network structure design, the MFF-CNN was based
on a densely-connected convolutional neural network design. By enhancing the feature reuse between network layers,
the parameter amount of the model was effectively reduced, and the problems of overfitting and gradient
disappearance of deep networks can be alleviated under the condition of small sample size. The experimental results
show that the MFF-CNN can attain the recognition rate of 91.6% on the image data set of the main bearing cap
collected in practice, and it displays good robustness in terms of the problem of uneven illumination of the parts’
images in actual production.

Key words: mechanical part recognition, convolutional neural network, fine-grained image classification, feature
fusion

中图分类号: