Convolutional Neural Network Based Natural Scene Text Detection And Application

As an important method of extracting visual data,natural scene text detection is an important research topic in the field of computer vision,and it has great application value in many different fields.Based on the convolutional neural network,the text detection algorithm for natural scene is studied and applied to the detection of pennant number.The main work and innovation of this thesis mainly includes the following several aspects:Firstly,the theoretical foundation of convolutional neural network is introduced.The implementation technology of deep learning algorithm based on convolution neural network as the main body are discussed in details,including the activation pool design,convolution function,function selection and training process etc.Inspired by Regional Proposal Network,the most effective object detection method,a fusion algorithm,the Holistic Vertical Proposal Network(HVRPN),based on the multi-scale deep learning architecture and the Regional Proposal Network is designed.The core of Regional Proposal Network is to use the anchor boxes of different sizes to generate candidate regions to fit the target area,and then classify and regress to detect the objects of different sizes.In order to further solve the problem of different size anchor points,the HVRPN method modifies the candidate region into a set of vertical candidate regions to break through the constraints of receptive field,thus meeting the requirements of different size text detection.At the same time,Holistically-nested networks(HED)is used to combine the low-level features and high-level features to improve the accuracy of text detection in different sizes.The method showed good effect in the ICDAR03 and ICDAR11 data.Finally,the natural scene text detection algorithm is applied to the pennant number detection task.According to the particularity of the pennant number,we optimize the detection model.And solve the problem of the of construction data shortage by transfer learning method.Then,according to the different characteristics of the pennant number,classification methods are applied to improve the accuracy of the algorithm.