Recognizing small-data Chinese invoices through pretrained models
Lu Zhong, Zidong Cui, Xiaoting Wang · 2023
Detection and recognition of text instances from invoice data is an important topic in optical character recognition (OCR). For conventional OCR models, large training datasets are usually required to achieve a high-quality learning performance. However, when only small training data are available, how to achieve a sufficiently low error rate is crucial for practical real-life applications. Here, we propose to combine the two neural networks, PSENet and TrOCR, to implement OCR for small invoice data in Chinese. Specifically, due to the fact that PSENet and TrOCR are pre-trained text detection and recognition models, they are expected to perform well even on small datasets. Simulation results show that our method can reach a low character error ratio (CER) of 0.0267 with a small training data of size 202.