A drawing retrieval model of mechanical parts based on YOLOv5 and a modified VGG network
Yuming Huang, Bin Ren, Jihe Feng, Kai-Wei Liu · 2022
How to retrieve similar drawings correctly from the existing drawing database is of great significance for companies. In this paper, we propose a drawing retrieval model for mechanical parts based on YOLOv5 and a modified VGG network (FEN-VGG). In this paper, the function of YOLOv5 is to detect stereograms and views in drawings and to classify the targets within the drawings into view, part stereogram, title & material block, and revision block. Then, we use a feature extraction network with pre-trained VGG network, FEN-VGG, to extract the features of stereograms and views respectively. And get the vectors composed of stereograms features and views features. After that, the similarity is calculated by the inner product with the vectors in the stereogram database and the view database respectively. Finally, the top 5 drawings with the sum score are output. The experiments show that the Top-1 accuracy of our model reaches 90% and the Top-5 accuracy reaches 95%. Top-1 accuracy and Top-5 accuracy improved by 36% and 31% compared to the previous respectively.