Abstract
With the development of economy and information technology, a large amount of invoice information has been produced. As one of the important components of the industrial Internet of Things, the recognition of invoice information is urgent to realize its intelligent recognition. Most invoice issuing units basically adopt traditional manual identification methods for the processing of invoices. As the number of invoices increases, problems such as low efficiency in identifying invoice information, error-prone, and difficulty in ensuring security frequently appear. In response to the above problems, this paper designs and implements an invoice information recognition system based on deep learning. The system first solves the problems of low image contrast and lack of image due to poor lighting or noise effects by image preprocessing methods such as image graying and normalization. Second, a target detection and invoice recognition method based on the combination of YOLOv3 + CRNN two models is proposed, and an end-to-end invoice information recognition model is obtained. Finally, the model is used to develop an invoice detection and recognition system based on deep learning. Experiments have verified that the system has the characteristics of high recognition accuracy and high efficiency, which can accurately identify invoice content information and reduce the loss of manpower and material resources.
Highlights
Foreign research on invoice recognition system originated in the 1960s and 1970s
Until 1986, Tsinghua University and other universities developed an invoice recognition system based on OCR technology, and Chinese OCR invoice recognition products came out [3]
Due to the low recognition rate of the early invoice system and insufficient productization, it has not been popularized in life
Summary
Foreign research on invoice recognition system originated in the 1960s and 1970s. But, most research is only on methods of invoice recognition and digital recognition. e concept of OCR (Optical Character Recognition) technology was first proposed by German scientist Tausheck in 1929. In the field of image recognition, researchers have further proposed a large-scale deep convolutional neural network, which reduces the error detection rate to 15.3% [7]. Deep learning has developed rapidly in image recognition, and target detection technology has been applied to text localization in natural scenes. The improved FastR-CNN algorithm in the research inputs the entire image into the convolutional neural network and maps the candidate frame on the feature map, avoiding repeated feature extraction and improving the training speed [11]. Is algorithm uses Faster R-CNN to extract features to improve the speed of target detection, which is very suitable for natural scenes with multiple anticounterfeit feature detection in invoices. In order to meet the requirements of efficiently identifying invoice data in engineering applications, this paper first uses the YOLOv3 algorithm for text target detection training. The two models are combined to obtain an end-to-end invoice recognition model, which is verified by the test set, and the recognition result is compared with the recognition result of the traditional OCR technology
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
