Lightweight Per-Field Ensemble of CRNN And Tesseract For CPU-Deployable Receipt Information Extraction

Abstract

Document information extraction from receipts is dominated by multimodal transformer architectures that achieve high accuracy at the cost of 150M+ parameters and GPU-dependent inference, rendering them impractical for edge and offline deployment. This paper presents a lightweight alternative that combines a 5.2M-parameter Convolutional Recurrent Neural Network with Tesseract OCR through a per-field confidence-gated ensemble. Unlike passive fusion strategies that fall back to a single engine, our approach actively routes each semantic field – company, address, date, and total – to the engine that empirically dominates it. Evaluated on the ICDAR-2019 SROIE benchmark, the ensemble achieves 0.602 average fuzzy similarity, outperforming Tesseract by 5.6% and the standalone CRNN by 41.3%. The CRNN dominates company-name extraction (0.672 vs. 0.594), where character-level CTC decoding handles stylized logos and irregular spacing, while Tesseract dominates numeric fields, where keyword anchoring and language modeling remain decisive. The ensemble never underperforms the best single engine on any field. Requiring only 21 MB of storage and 0.12 s per image on CPU, the framework is suitable for offline MSME deployment, closing the gap between lightweight inference and hybrid fusion accuracy in document intelligence.