Image Format P&ID Recognition Technique Using an Oriented Symbol-Text Integrated Recognition Model

Existing deep learning-based research on P&ID recognition utilizes separated detection model to recognize symbols and text within image format P&IDs. In conventional studies, since the text detection and recognition models are independent, the features extracted by the text detection model cannot be utilized by the text recognition model. Therefore, text recognition models that use text images as input require a feature extraction module, which increases both the model size and recognition time. Additionally, since text detection and recognition models are trained independently, semantic information cannot be effectively exchanged between them, limiting improvements in text detection and recognition accuracy. To address these limitations, we propose an integrated model that combines the symbol-text detection module and the text recognition module, using an oriented object detection module capable of recognizing rotated objects as the symbol-text detection component. By leveraging features from text regions detected by the orientation-aware object detection module, the proposed model enables fast text recognition with a compact structure, reducing text recognition time per drawing by 1.595 seconds. The joint training of the orientation-aware symbol-text detection and text recognition modules improved text detection and recognition outcomes, achieving a 2.81% improvement in text recognition performance based on F1-score.