Invoice Image Text Detection And Recognition Based On Deep Learning

With the development of deep learning,many problems in scene text detection and recognition have been solved.Invoice recognition has an urgent practical need.The research of invoice recognition has great application value.But at present,there are few public research results on invoice recognition.Most of them are template-based methods,which have relatively low performance and low robustness.The advantages of deep learning have not been exploited.Because of the flexibility of invoice itself,the effect of typing,and the environment of shooting,invoice images often have the characteristics of deformation,unclear text,incomplete and poor image quality.These characteristics are the key factors affecting the performance of invoice recognition,and are also the research difficulties of invoice recognition.These imaging features of invoices have many similarities with the problems faced by scene text detection and recognition,and the text on the invoice is presented differently from the scene text.In response to these difficulties,this paper combines the latest research results of scene text detection and recognition.On the basis of end-to-end scene text detection and recognition network FOTS,we replace the text detection method of FOTS network,and change the method based on corner point text detection and position sensitive area segmentation to improve the invoice image recognition problem.The enhanced network is robust to detecting multi-directional and multi-scale text,and finally designs an end-to-end network structure.In addition,this paper also proposes the feasibility of using program to synthesize invoice image,and uses the synthesized invoice image data set to train deep learning model in the experiment.The experimental results show that on the invoice image data set,the F-measure of text detection reaches 92.16%,the overall recognition precision of valid invoice information reaches 96.56%,and the average time of network processing single invoice is 264 ms.This shows that the scheme based on deep learning can effectively improve the precision and speed of invoice recognition.