Research On Application Of Character Recognition Based On ResNet Network

In recent years,computer vision has developed rapidly,and related applications have also entered into people's daily lives.There are still many problems to be solved in recognizing text in natural scene images.Words in natural images usually have irregular shapes,which are caused by perspective distortions,curved character placement,etc.In some scenes,there are curved and deformed texts,resulting in a low recognition rate for some recognition methods.In order to improve the text recognition rate,I started from the feature extraction network of the text recognition network and replaced the ordinary CNN network with the ResNet network.Because the ResNet network is deeper,it solves the situation of network degradation.If the CNN convolutional layer is continuously deepened,it may be A gradient explosion occurs.At the same time,digitizing paper documents can be conveniently stored in the database to save and complete the data query.With this as the background,the research on the text recognition system was started,and a simple system for human-computer interaction was built.This article first introduces the relevant background of the topic and the development status of word recognition.In a chapter,I will explain the improvements made in the feature extraction part of the CRNN text recognition network.It mainly compares the results of the feature extraction used by the ordinary CNN network and the ResNet network to identify the text.Through the comparison of the experimental results,the effect of the ResNet network extraction is found.better.Then based on the RARE network structure,ResNet network was used to replace the original feature extraction network,and each part of the text recognition network was described in detail.RARE text recognition network used STN network to complete text correction for irregular text,and the network model was End-to-end training,serialize the text in the image,use the Attention mechanism to complete the decoding process,and design experiments to complete the verification,and found that the improved network model recognition after improved feature extraction network is better.Replace the feature extraction network in the CRNN network and RARE network with the ResNet network.The improved network performs better on the IIIT5 k,SVT,ICDAR2013 and ICDAR2015 data sets.Use the improved RARE network to train the text recognition model of the Chinese data set.For text recognition system.Finally,we introduce the design of the interactive system for text recognition,the CTPN network for text area detection,some basic functions displayed on the front-end page,such as image upload and preview,and the display of recognition results.