The Research And Implementation Of Text Recognition And Search Based On Deep Learning

With the popularization of portable photographic devices such as smart phones,more and more people are using images as the preferred information carrier.The continuous innovation of computer vision technology,new artificial intelligence application such as driverless cars,human-computer interaction and intelligent security are also increasing.In the information age,how to make the computer understand image information accurately is attracting more and more attention.On the one hand,text elements are rich in semantic information,recognition of natural scene text information can help computers understand image information more accurately.On the other hand,traditional text recognition technology is only suitable for printed document recognition with same character size,uniform fonts and fixed text intervals,such as various documents,bills,documents,etc.However,because of the complex background,variable font size and various noises,the traditional text recognition methods are difficult to recognize text in the natural scene accurately and effectively.The continuous expansion of image data sets higher requirements for the classification,coding,storage,retrieval and maintenance of image data,especially for the retrieval of image data.It is also a new application direction to construct a retrieval application that retrieves the target image by retrieving text information.In recent years,with the continuous development of deep learning technology,it is different from the feature extraction based on manual experience.Depth learning can acquire data features spontaneously by learning massive training data,especially for pattern recognition such as object recognition and speech recognition.This paper realizes the detection of text area in image based on CTPN algorithm,and realizes the recognition of text information in text area through DenseNet network and CTC algorithm.Finally,it realizes the application of image retrieval by combining the result of text recognition and the reading and writing of image Exif information.The main contents of this paper are summarized as follows:(1)This paper realizes the CTPN text detection model.CTPN algorithm combines the characteristics of text objects,uses the vertical anchor mechanism to detect a small rectangular box,and realizes the text line detection by merging rectangular boxes.Considering that the continuous context information in the text line has a good reference for detection,CTPN also improves the detection effect by adding BLSTM.In order to verify the extensiveness of the algorithm,100 labeled images were tested.The experimental results show that the average performance of the algorithm can reach 87.6%when the evaluation is DetEval,and it can successfully realize text detection in natural scenes.(2)This paper implements a text recognition model based on DenseNet+CTC.DenseNet network ensures that the maximum information between network layers by connecting the feature maps of all layers.DenseNet uses narrower network structure and fewer parameters.The concatenate mode also makes the transmission of features and gradients more effective.After getting the features through DenseNet network,CTC loss function is used to solve the network training problem in the case of input-output misalignment.In this paper,we randomly select the continuous characters in Chinese corpus,and get 364,000 picture data through a series of data enhancement.Then we divide the training set and test set by 99:1.Finally,we further optimize the text recognition model through the test results on the labeled data.The experimental results show that the accuracy of the optimized model can reach 70.6%when the text similarity is 0.8.(3)This paper implements an image retrieval application based on text recognition,which uses Flask Web framework.Firstly,the text information in the image is obtained through the text recognition model,and then the recognition result is written into the Exif information of the image with the help of Piexif expansion package.The background of the Web reads the Exif information of the image and matches it with the search words,obtains the target image and displays it on the front page of the Web.This paper realizes the application of image retrieval which retrieves the target image through text.