Exploring the Potential of OCR Integration for Object Detection in Invoices
Andrei-Ştefan Bulzan, Cosmin Cernăzanu-Glăvan, Marius-George Marcu · 2023
This paper investigates the impact of incorporating Optical Character Recognition (OCR) information into object detection models for extracting key information fields from invoices. We propose a method that adds a fourth channel to the input images, representing text presence, derived from two OCR models. Our experiments show that while larger models do not benefit from the additional text localization information, smaller models exhibit significant accuracy improvements and accelerated learning. In particular, we observe a substantially higher mean average precision (mAP) by epoch 20 out of 100 when including the fourth OCR information channel. This research demonstrates the potential benefits of incorporating OCR information into object detection models, particularly for smaller models with limited resources, by enhancing not only accuracy but also the speed of convergence during training.