Research And Implementation Of Financial Invoice Intelligent Recognition Based On Deep Learning

With the rapid development of China's economy and science and technology,the improvement and identification of invoice content is also in continuous improvement.The processing and filing of invoices are almost the same work for the financial personnel of enterprises or institutions.This kind of highly repetitive work greatly increases the labor cost and the time cost of financial personnel.With the development of neural network research in recent years,the application of deep learning in image text recognition has become a hot research direction.It is very important to recognize the words in the invoice image information for reducing the cost of manpower and material resources.In this paper,the common value-added tax invoice in business activities is taken as the object to study the intelligent recognition method of invoice.The main work of this paper is as followsFirst,a scanner is used to capture high-quality invoice images.In order to improve the accuracy of invoice text detection and recognition,this paper improves Hough transform to detect the tilt angle of invoice image by taking the long horizontal line in the invoice layout as the target.After that,the seal of invoice image is removed to reduce the interference of text detection and recognition.Secondly,the invoice text detection module in this paper is based on the YOLOv3 detection algorithm.According to the characteristics of the invoice data set,the large-scale feature detection branch in the original YOLOv3 model is deleted,and a more accurate feature detection branch is added to enhance its detection effect on small targets;K-means++ clustering algorithm is used to re cluster the data set to get nine new anchor boxes;In order to make the model run on the equipment with weak computing power,this paper prunes and compresses the yolov3 model,reduces the model parameters to 20% of the original,and improves the running speed of the model.In the text recognition module,this paper uses the end-to-end deep learning algorithm based on convolutional recurrent neural network to recognize the extracted invoice text field.The algorithm can obtain a wide range of context information and flexibly identify text sequences with variable length.For the text image without serious noise,it can achieve error free level.The experimental results show that the convolutional recurrent neural network can effectively identify the fields in the invoice.Finally,based on the seriously defaced invoice which can not be effectively identified,this paper identifies the invoice QR code by using the required fields for the VAT invoice inspection platform.In the positioning module,we use the yolov3 deep learning detection algorithm to enhance the two-dimensional code data set,and use k-means++ to cluster the data set again.The final model achieves real-time detection while ensuring the detection accuracy.