Efficient Information Recognition for Machine-printed Invoices

Automatic recognition of information in machine-printed invoice images is significant to enterprises in saving labor. However, most invoice information identification methods rely on high-quality invoice images, and collecting these images requires the purchase of professional and expensive image acquisition equipment. In this paper, we propose a novel invoice recognition framework capable of identifying lower-quality invoice images captured by devices such as mobile phones. we present an image preprocessing algorithm to correct distorted invoice images based on perspective projection. Then, accurate content Key-Value (KV) structures are extracted by U-net and template matching combined with the coordinate fine-tuning method, and the Robust text recognizer with Automatic Rectification (RARE) is adapted to recognize characters. To further improve the proposed model in small data scenarios, we adapt the Generative Adversarial Networks (GAN) to extend training samples and improve the RARE by active learning method. Extensive experiments are conducted on both machine-printed invoice images collected from actual scenarios and the open-source ICDAR dataset. The comparison results demonstrate the superiority of the proposed method on both sufficient and small data scenarios.