Intelligent automation of invoice parsing using computer vision techniques

Manual parsing of invoices is a tedious, arduous and error-prone task. Due to the academic and business importance of this problem, it has attracted the attention of machine learning enthusiasts. There are several complexities and challenges in the automated parsing of invoices. Some of them include a paucity of useful datasets, eclectic template formats, and poor performance of algorithms in real life scenarios. This problem can be solved by the automatic traversal of the invoices by object detection algorithms such as YOLO, SSD and R-CNN. These state-of-the-art algorithms will be trained to detect various fields or entities present in an invoice. In this paper, a dataset of 315 invoices has been generated using web testing tools. The dataset has been annotated for eight entities: billing address, shipping address, invoice date, invoice number, product name, price, quantity, and total amount. The text boxes detected by the models is converted to machine encoded text, using text extraction methods such as Optical Character Recognition (OCR). Hyperparameter tuning has been performed to improve model accuracy. The models have been evaluated on myriad metrics such as mean Average Precision (mAP), common objects in context (COCO) evaluation metrics and total loss during training and validation. The loss vs iteration graph has been visualized using Tensorboard. A front-end application encapsulates all the functions of the research paper and allows testing of various models.