Nowadays, the level of invoice traffic between companies has reached enormous levels. Invoices are crucial financial documents for companies, and they need to extract this information from these documents to access and control them quickly when necessary. While electronic invoices can be easily transferred to the company’s ERP system with the help of integrators, information from printed invoices must be entered into the ERP system. Information entry is generally performed manually by company employees, so the probability of error is high. The automatic recognition of information in printed invoices will reduce the possibility of error. It will also save time and money by reducing workforce requirements. This study proposes a deep learning-based solution for detecting fields in image invoices that are in high demand among businesses. The system offers an end-to-end solution, which includes a novel method for generating synthetic invoices and automatic labeling. Three invoice templates were used to evaluate the usability of the system and an adaptive fine-tuning-based solution is proposed for newly coming invoice templates. Furthermore, 6 different object detection models were compared to find the most suitable one for our problem. The system was also tested with 1022 real invoice images that were manually labeled to test real-world usage. The results indicated that the fine-tuned model achieved an accuracy that was 8.4% higher than the baseline models. In tests performed on CPU, TOOD and Cascade-RCNN models were the most successful algorithms, while YOLOv5 was the fastest running algorithm. Depending on the priority of the needs, both algorithms can be preferred for real-time usage in the detection of invoice fields. The synthetic invoice generation code is available at https://github.com/SCU-CENG/Invoice-Generation.

如今，公司之间的发票流量已经达到了巨大的水平。发票是公司至关重要的财务文件，他们需要从这些文件中提取信息，以便在必要时快速访问和控制它们。虽然电子发票可以在集成商的帮助下轻松传输到公司的 ERP 系统，但打印发票的信息必须输入 ERP 系统。信息录入一般由公司员工手工完成，因此出错的概率较高。自动识别打印发票中的信息将减少出错的可能性。它还将通过减少劳动力需求来节省时间和金钱。本研究提出了一种基于深度学习的解决方案，用于检测企业需求量很大的图像发票中的字段。该系统提供了端到端解决方案，其中包括生成合成发票和自动标签的新颖方法。使用三个发票模板来评估系统的可用性，并为新推出的发票模板提出了基于自适应微调的解决方案。此外，还比较了 6 种不同的目标检测模型，以找到最适合我们问题的模型。该系统还使用 1022 个真实发票图像进行了测试，这些图像经过手动标记以测试实际使用情况。结果表明，微调模型的准确度比基线模型高 8.4%。在CPU上进行的测试中，TOOD和Cascade-RCNN模型是最成功的算法，而YOLOv5是运行速度最快的算法。根据需求的优先级，这两种算法都可以优先用于实时检测发票字段。合成发票生成代码可从 https://github.com/SCU-CENG/Invoice-Generation 获取。