Improved P&ID Symbol Detection Algorithm Based on YOLOv5 Network

The existing object detection algorithms have the problems of poor detection performance on the piping and instrumentation diagram(P&ID) symbol dataset. Therefore, this paper proposes a symbol detection algorithm in P &ID images using improved network YOLOv5. Firstly, due to the limited number of samples in the symbol dataset and the class imbalance problem, this paper proposes a Random Splicing algorithm suitable for P &ID images. It can selectively augment one or more classes of symbols in P&ID images. Then, to address the problem of subtle differences and difficulty in dis-tinguishing between symbols, this paper introduces the CBAM and improves its “series” structure to a “parallel” structure, enhancing the information fusion between channel feature maps and spatial feature maps through concatenation and channel shuffle, after using a
$1\times 1$
convolution to achieve channel alignment, resulting in the PCSC_CBAM (Parallel Concat Shuffle Convolution_CBAM). Finally, the YOLOv5+PCSC_CBAM model is obtained by introducing PCSC_CBAM into the Neck of YOLOv5, which can effectively detect symbols in P&ID images and reduce the possibility of false negatives and false positives. The experimental results show that under only 1.97M parameters added, the YOLOv5+PCSC_CBAM model achieves high precision, recall, and mean Average Precision of symbol detection, reaching 95.6%, 95.0%, and 97.2%, respectively.