UDMtext: a Unified Deep Mutual Learning-Based Framework for Text Detection in Complex Documents

With the increasing demand for digital document processing and analysis across various industries, the need for efficient and automated handling of complex documents has become more urgent. Existing Optical Character Recognition (OCR) methods face challenges such as inaccurate text localization and poor robustness when dealing with documents containing complex content. To address these challenges, this paper proposes a complex document text detection framework based on Unified Deep Mutual Learning (UDMText). The framework integrates a teacher model with two lightweight student models in a collaborative learning mechanism. Through deep mutual learning <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[1]</sup>, the student models further optimize feature representation and prediction capabilities based on knowledge distillation, thereby enhancing their understanding of multimodal features. In addition, we design a hybrid loss function that effectively balances global feature alignment with local detail optimization, improving the model's detection accuracy and convergence speed. On the ICDAR 2023 dataset, the UDMText framework demonstrates significant performance improvement in end-to-end text detection and recognition tasks. Experimental results show that UDMText achieves an accuracy of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 2. 9 1 \%}$</tex> and a recall rate of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{6 2. 3 1 \%}$</tex>, while exhibiting outstanding robustness in complex document layouts, validating the effectiveness of the framework in complex document processing. Our work provides a new solution for the application of OCR technology in complex document scenarios and offers theoretical and technical support for future research.