Research On The Methods Of Chinese Characters Detection And Recognition In Natural Scene Images

With the development of new technologies such as intelligent transportation systems,visual aids and autonomous driving,capturing scene images through cameras and using computers to analyze and understand the contents of scene images has become a research hotspot in academic and industrial circles in recent years.Compared with the complicated background,the text line in the scene image carries more useful information and is an effective clue to describe the content of the scene.Therefore,the text location and recognition technology in natural scene images has become an important research topic in the field of computer vision and pattern recognition.However,at present,the research at home and abroad mainly focuses on the recognition of English characters in natural scenes,while the research on Chinese characters is still rare.And because Chinese characters have the characteristics of complex structure,huge character set,high similarity of characters and diverse styles of fonts,it is more difficult to locate and recognize Chinese characters than English characters in scene images.This paper mainly focuses on the localization and recognition algorithms of Chinese characters in natural scenes.Based on the algorithm of maximal stable extreme regions(MSER),a method to extract regions of interest by color clustering is proposed by simulating human visual mechanism.At the same time,false alarms are filtered out by further combining the stroke width,corner features and some other heuristic rules based on the geometric features of Chinese characters.Finally,a deep convolutional neural network is used as a binary classifier to distinguish Chinese characters from backgrounds.The background information with high similarity to Chinese characters is further removed in this way,so that the text regions in the scene image can be obtained more accurately.In the process of designing and training the convolutional neural networks,a variety of innovative strategies have been introduced such as data augmentation,batch normalization and fine-tuning the model by combining two optimizers,namely,stochastic gradient descent(SGD)and adaptive moment estimation(Adam).Besides,the network structure and loss function are improved according to the binary classification situation.Through the above methods,the goal of using a limited amount of data to improve the classification effect of the network as much as possible while shortening the training time is achieved.The validity of the method proposed in this paper is verified by experiments on a self-built scene image database,and the recall rate and accuracy of the texts localization in the scene images are higher than the existing methods.This paper also carries on a deep research on the Chinese character recognition technology.A database which contains a total of 3755 categories of Chinese characters and 22 kinds of fonts including printed and approximate handwritten styles is established.A deep convolutional neural network model is proposed for processing unscreened images with printed and handwritten Chinese characters.Regardless of the additional layers,such as batch normalization and dropout layers,the network mainly consists of three convolutional layers,two pooling layers,one fully connected layer and a Softmax regression layer.By introducing a variety of innovative training methods including data augmentation,batch normalization and using two kinds of optimizers,the generalization ability and recognition accuracy of the model are effectively improved and the accuracy rate reaches 98.336% on the self-built database.