Design And Implementation Of Medical Invoice Information Extraction System Based On Deep Learning

In recent years,deep learning has become very popular in the academic community,especially in the field of computer vision.More and more deep learning based vision solutions are proposed and significantly outperform traditional methods in the past.This makes it possible to use deep learning to complete the transformation of traditional business.Currently,there is a great demand for the detection and recognition of medical invoices in some insurance enterprises.In the past,some traditional methods based on template have some difficulties(such as low efficiency,low precision,and poor robustness)when dealing with complex images.Therefore,at present,most enterprises still audit and reimburse bills manually.There is a serious waste of time and cost.With the development of text detection and recognition technology in natural scene,it becomes a feasible solution to solve the problem of invoice information location and recognition.Most medical invoices are usually collected by camera.The text in the image often has multi direction,deformation,complex background and so on.These features are very similar to the text in the natural scene.Besides,due to the printing quality and the interference of the seal,it is more difficult to detect and identify the invoice.Aiming at these difficulties and combining with the existing text detection and recognition technology,this thesis proposes a set of solutions to realize the location and recognition of medical invoice by using the deep learning technology.Corner detection and position sensitive segmentation are used to locate the invoice information.In addition,a feature enhancement module is added to the network to extract more robust multi-scale features and further improve the detection results.In the recognition phase,this thesis adopts an end-to-end text recognition method based on context information fusion.In the part of feature extraction,convolution attention module is introduced to help the network select features adaptively.In addition,in order to enhance the learning of Chinese semantics,this thesis conducts a large-scale data synthesis based on the collected general Chinese corpus and invoice text corpus.The validity and practicability of the method proposed in this thesis are proved by testing on two kinds of real datasets,table and machine printed invoice.