CNN-based English Text Detection And Recognition System For Natural Scene Images

With the popularization and development of the Internet,our life is full of images.These image data contain a large number of information important to users,including image information and text information.Text information may include the most important information of the image,which can be used in the work of paper document preservation,identification of identification,identification of license plate number,identification of applicable card,and instant video translation.In the past,people used to extract image information by manual marking.But in the face of a large number of images,this method is obviously not desirable.Then,people have made great achievements in the steady extraction of scanned text images with the development of sleep OCR technology.Text information in natural images has great research value,but there is no such application in the market.The research object is images of natural scenes,containing text information in complex backgrounds.A set of algorithms using convolutional neural networks is designed through two steps of text localization and text recognition.Based on this algorithm,a web system is developed.The specific work is as follows:1.An effective image text recognition algorithm is proposed.First of all,using stroke width transformation(SWT)algorithm to supplement the data set--because CNNs can have millions of trainable parameters,I need a lot of training data to minimize generalization,and mining helps to expand available data cheaply;I grab images from the Internet,automatically generate word level and character level boundary box annotations,when only word level boundary box annotations are provided A separate method is used to automatically generate character level bounding box annotations.Then,a scheme of feature learning based on convolutional neural network is proposed,which uses four layers of CNN for feature learning.Finally,text / background classifier,case insensitive classifier,case sensitive classifier and binary character classifier are trained.It is worth noting that these four CNN share the first two layers of case insensitive classifier.Finally,convolution neural network model is used to locate and recognize text.The probability map obtained by text / background classifier is used for thresholding to find local areas with high probability,and RLSA is used to connect these areas to the written lines;the text lines are divided into words,and Otsu algorithm is used to separate foreground characters from the background.If the character spacing is less than the average text spacing,RLSA is used to connect these areas again to get the word areas 。Using case insensitive classifier,case sensitive classifier and binary character classifier to recognize characters,we synthesize the results of three classifiers and get the most likely characters.2.Analyze and design the requirements of the natural scene image text recognition system.According to the different use cases,analyze the basic functions and data models of the system.In the design of the system architecture,functions and database,describe the specific operation and implementation of the system.In the implementation of thesystem,firstly,the network model is trained by using the locally downloaded data set,and the important parameter files are saved to identify the pictures uploaded by users.Then,a web system is built to realize a series of functions of image recognition.This system provides a complete solution from user management,image recognition to image management,so that users can manage personal image resources more conveniently.