TrOCR-driven seal instrument detection and recognition for cognitive robotic applications
Xuan Jin, Sheng Wang, Miaomiao Zhang, Guoteng Xu, Bingqi Hu, Hanlin Tang · Cognitive Robotics · 2025
Seal recognition, as a fundamental perception capability, is crucial for enabling cognitive robotic systems to autonomously interact with and understand physical documents in intelligent office and archival environments. While Transformer based optical character recognition (OCR) methods have recently achieved remarkable progress, the recognition of curved and degraded seal text remains a significant challenge. Traditional approaches often rely on cumbersome pipelines with limited robustness, which hampers their integration into robotic cognitive platforms. To address these issues, this paper proposes a novel perception framework that integrates the YOLO-based detection module with the TrOCR recognition model for seal content analysis. The framework enhances robotic perception through three core mechanisms: precise spatial localization, adaptive noise suppression, and efficient curved-text recognition. Experimental results demonstrate that the proposed approach achieves 94.8% accuracy in bent seal text recognition tasks, validating its effectiveness in complex, real-world scenarios. These findings highlight the potential of the method to serve as a reliable perception module within cognitive robotic systems for document understanding and autonomous decision-making.