Key points are not available for this paper at this time.
Efficient extraction of critical information from receipt is essential for automating financial processes and supporting timely decision-making in businesses. However, this process faces significant challenges, starting with variations in the quality of scanned receipt images due to differences in scanning equipment, followed by the complexity of diverse receipt formats, and further complicated by handwritten elements and noise, making accurate extraction particularly difficult. Therefore, to address these issues, we propose a model framework called DocExtractNet, based on LayoutLMv3, designed for extracting key information from receipt. Firstly, we introduce the ImageEnhance method to process image modality features, enhancing image clarity and significantly improving recognition accuracy for low-quality images. Then, we implement the PrecisionHints strategy to supplement missing key–value pairs in the text modality, improving data integrity and the model’s overall performance. Furthermore, we apply the CrossModalFusion method to combine both image and text features, allowing the model to better understand and extract receipt information. The experimental results on the Finance-Receipts, FUNSD, and CORD datasets show that DocExtractNet significantly improves F1 scores compared to other models, with F1 scores reaching 97.07% for Finance-Receipts, 91.80% for FUNSD, and 97.38% for CORD, highlighting its superior performance in receipt information extraction. • Introduced a novel model for key information extraction from invoices. • Enhanced multimodal information extraction by integrating data from different modalities. • Achieved significant accuracy improvements in complex invoice datasets through innovative fine-tuning strategies. • Optimized model structure and parameters, leading to increased processing efficiency for invoice information extraction tasks.
Yan et al. (Mon,) studied this question.