Text Detection In Natural Scene Images

With the continuous development of mobile Internet technology,the way of information dissemination is no longer limited to simple text forms.More and more videos and images are flooding in the Internet today.Therefore,how to obtain effective information from video and images is becoming more and more important,which is of great help for us to understand,storage and classification of images well.The text in the images is usually a supplementary description of the video and images,which plays a decisive role in the understanding of the video and images.Deep learning technology has developed rapidly in recent years,especially in the field of object detection in images.T his method usually extracts the deep image features in the image by means of the deep convolutional neural network,and classifies the images,which does not require manual intervention to set the image features,and can spontaneously learn from a large amount of input data set.This efficient learning method has also greatly promoted the development of the field of obj ect detection in images.Based on the research results which has achieved in the field of object detection already,this thesis applies the current popular object detection network model Faster Region-Convolutional Neural Networks(Faster R-CNN).And considering the specificity of text detection tasks in images,we have improved the Region Proposal Network(RPN)to implement a new network structure that can realize rectangular stripe text area detection of equal width.And then,we proposed a text line construction algorithm using linear regression theory,which will help us to recombine a series of equal-span rectangles that detected from the text detection network into a continuous and complete text line area.In addition,this thesis proposed a morphological operation algorithm for specific experimental environment,which can process the constructed text line area and extract the accurate boundary contour of the text region in images by means of morphological algorithms in traditional image processing.The algorithm can ensure acquire the boundary contour of the text area accurately,and meanwhile,remove the isolated point noise area in the text detection result,which greatly improves the accuracy of image text area detection.In the end,we experiment and verify the text detection network model through multiple public data sets of text detection.our text detection network performs well in the experiments,verifying the effectiveness of our network model.