Reproduce “An Efficient Deep Learning-Based Approach to Automating Invoice Document Validation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Aziz Amari, Mariem Makni, Wissal Fnaich, Akram Lahmar, Fedi Koubaa, Oumayma Charrad, Mohamed Ali Zormati, Rabaa Youssef Douss
A practical invoice-validation pipeline combines OCR and layout-aware field extraction with object detection for stamps and signatures. On a small collection of real, imperfect financial documents, LayoutLMv3 achieved the best field-recognition F1-score among the tested layout models (0.56), while RetinaNet led stamp and signature detection with [email protected] of 83.96.
Key insights
- LayoutLMv3 provided the strongest field-extraction result, but its advantage over LiLT was limited and its validation curves indicated overfitting; LiLT was approximately twice as fast in inference.
- The OCR comparison produced conflicting signals: Tesseract yielded higher directly measured F1-score than Azure OCR on the invoice data (0.81 versus 0.56), although Azure OCR was selected for the final pipeline because it generated more text instances.
- Authenticity checks require a separate visual-detection component: RetinaNet offered the best accuracy, whereas SSD was fastest but substantially less precise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.