Detection And Recognition Of Medical Bill Based On Deep Learning

In recent years,with the rapid development of deep learning,text recognition under natural scenes has made great progress,such as license plate recognition.However,bill texts(especially medical bills)in natural scenes have problems such as dense text and misprinting compared to other texts.Therefore,bill text recognition has always been a research difficulty in the field of computer vision.Based on the above problems,this paper proposes a complete set of medical bill text recognition solutions.This set of solutions mainly includes two parts: text detection and text recognition.In the text detection task,due to the dense text of medical bills and the existence of a large amount of text bending,this paper uses the progressive scale expansion network(PSENet)as the basis to carry out the text detection research.In addition,in view of the multi-category text(including printed and machine-printed text)in medical bills,this article quotes the multi-classification improvement scheme proposed by the predecessors based on PSENet.This scheme will also classify text instances while detecting the text,which is a follow-up Convenient processing.In addition,this paper has made a series of improvements to PSENet:(1)The feature extraction backbone network is changed from Res Net to a lightweight network Mobile Net V2,which greatly improves the efficiency of model calculations;(2)Repair the broken text that appears in the detection through the morphological closed operation.Finally,while reducing the size of the network model by14 times,the F1 score of the test result was increased from 76.3% to 85.8%.In the text recognition task,this paper uses the currently popular CRNN+CTC model.Unlike the text in other natural scenes,the text in a medical bill contains a large number of fixed fields,such as detailed items in a printed template.Therefore,this paper uses a transcription scheme that combines no dictionary transcription and dictionary transcription to output the final recognition results,and finally increases the accuracy of field recognition in the bill from 89.8% to 91.2%.Finally,this article trains and verifies the feasibility of the whole scheme on the labeled Chinese medical bill data set.At the same time,the program was compared and tested with the commercial solutions of Baidu and Face++,which proved that this solution has certain advantages in the text recognition of medical bills.