Tag classification and detection for piping and instrumentation diagrams

Symbol detection methods for Piping and Instrumentation Diagram (P &ID) have been continuously developed over the past few decades, evolving from traditional methods to convolutional neural networks (CNN). This study aims to compare the performance of tag classification and detection in terms of both accuracy and time consumption between traditional methods, a designed-from-scratch CNN, and ResNet50 transfer learning using the same dataset. The results show that ResNet50 transfer learning achieves the highest F1 at 85.9%, but takes the longest execution time at 175.7 s per diagram. Meanwhile, most of the errors in the traditional method and the designed-from-scratch CNN were false positives in tag detection for the diagram description on the right pane. However, after applying a rule to crop the diagram picture before classification and detection, the designed-from-scratch CNN exhibits the best performance, achieving the highest F1 at 89.6%, with a running time of 118.4 s per diagram, which is comparable to the traditional method.