Design And Implementation Of A Natural Scene Text Recognition System Based On Deep Learning Algorithm

With the rapid development of the popularity of smart mobile phone, and theInternet through the mobile phone, tablet computer and mobile terminal cameraacquisition, processing and sharing of information has become a very popular way oflife. Based on the camera (Camera-based) application pays more attention to thenatural scene understanding. In general, text and other objects in the scene are oftenexist at the same time, the user will pay more attention to the text in natural scenes,so how to accurately and fast recognition, text in natural scenes, there will be morein-depth understanding of the user’s intention and the theme of photography.However, there is little research on natural scene image text recognition and relatedalgorithm is also very immature, we need further research and exploration.Based on the analysis of the progress on various designs of word recognition ofnatural scenes in the picture, and combined with the deep learning algorithm, Iimplemented a perfect natural scene image character recognition system.The system this paper studies is a natural scene text recognition system, it canautomatically identify text information in natural scenes picture. The system consistsof image preprocessing module, unsupervised feature learning module, CNNcharacter detection module, CNN character classification module and a text linedetection module. Image preprocessing is the comparison of the normalized image,principal component analysis and whitening operation so as to carry outunsupervised feature learning, learning to feature dictionary of characters. Characterdetection and character recognition using convolutional neural network of twodifferent structure. Convolutional neural network character detection module in thefirst layer convolution kernel is obtained by unsupervised feature learning methods.The output of the neural network is connected to the support vector machine linear,as judged by whether the block of pixels for the character of the classifier. Thecharacter recognition module structure and character detection module is similar, the only difference is that the connection is the multilayer perceptron. For a picture, firstthrough the CNN character detection module is calculated for each pixel block score,then use non maxima suppression algorithm to locate the text in image. At the end ofthe characters in the text line position calculating character classification score usingCNN character recognition module, and search for the best word in the glossary.Natural scene image character recognition system is tested to meet userfunctional requirements. To assess the performance of the system, the realization ofthe recognition system is tested on the test data set, the results show that theperformance can meet the expected goals: training accuracy of character detectionwas94.47%, validation accuracy is93.47%; training accuracy of characterrecognition is98.87%, validation accuracy of74.22%.