Scene Numbers Recognition Based On Convolutional Neural Network

As the human society moves into the era of Big Date,more and more multimedia data springs up in the Internet.In the face of numerous image data,people are eager to identify and process the multimedia data with computer automatically,which contributes to the development of computer vision.Within this field,how to extract text information from the complicated background of a picture is always the difficult and hot issue.In recent years,the application of neural network allows breakthroughs in the field of computer vision.Compared to traditional way of extracting image features manually,neural network can automatically extract high-level features,which is quite important for solving problems like processing natural scene and thus make this method advantageous.Moreover,convolutional neural network possesses its characteristics which allow it to avoid exponential growth while processing high-dimensional data like images.Therefore,it gradually becomes a trend to apply convolutional neural network to text identification of images.In this context,this thesis divides the digital recognition of natural scene into two tasks,namely,character detection and character recognition.Firstly,convolutional neural network is applied to locate the charater area.After obtaining the exact position of the character area,the recursive neural network is used to identify the string contained in the region.By analyzing and comparing object detection and character recognition of natural scence,this thesis applys Faster-RCNN,a mainstream framework of object detection,to the charater detection and regards the string as an object.In the application of Faster-RCNN,the output of the framework,network size,anchor ratio and IOU threshold,and several aspects are optimized in accordance with character recognition.As for character recognition,a network which integrates convolutional network and recursive network is used,with convolutional network extracting features and recursive network generating final character sequence.Both networks are trained to make up a complete identification system and verification is done with several public data sets.Finally,the accuracy of character detection is apparently higher than other methods.