Application Of Text Detection Based On Semantic Segmentation In Receipt Optical Character Recognition

Text,as a carrier of information,often contains a wealth of important information that people need.scene text recognition is to detect and recognize text examples in natural scenes,such as license plates,street signs and billboards.Therefore,scene text recognition has a broad application prospect in daily life.The research content of this article is how to accurately extract the text information in the medical bill image,so as to ensure the accuracy of subsequent text recognition.Because that entering the data in the medical bill into computer is costly and inefficient,receipt entry task is heavy and intensive.So that make the entry personnel feel too tired to make a mistake in work.However,if we can use the scene text detection algorithm to complete the task of entering bill information through the machine,the burden of manpower can be reduced and the accuracy of the machine will not be reduced due to fatigue.In order to solve the difficulties of text detection in medical bill images,this paper proposed a new type of solution – Multi-class Progressive Scale Expansion Network(MPSENet),based on the most advanced neural network in the field of scene text detection – Progressive Scale Expansion Network(PSENet).The biggest difference between MPSENet and PSENet is that MPSENet changed the output type of PSENet from singleclass output to multi-class output,which means MPSENet can not only detect the text area,but also classify the text according to the color and font characteristics of the text.The reason why this paper design network like this is that text instances in the medical bill can be divided into several categories(including printed text,seal text,etc.),and different types of text instances will cover each other.Therefore,the above problems can be solved well by detecting these characters separately,and we can also process the text instances separately in the subsequent processing.This can reduce the workload and difficulty of subsequent processing,and improve the recognition accuracy.During the experiment,this paper prepared two Chinese scene text detection datasets to verify the feasibility of the network proposed in this paper.These two datasets are the synthetic scene text detection dataset and the medical bill dataset that we collected and labeled ourselves.Finally,the experimental result shows that MPSENet achieves a F-measure of 76.00% in the medical bill dataset.What's more,we compared the test results of commercial solutions of MPSENet,Baidu cloud and Face++,which proves that the detection result of MPSENet is better than that of these commercial solutions to a certain extent.