Research On Text Detection Technology In Natural Scene

As one of the research hotspots in the field of target detection,text detection technology in natural scenes has a wide application prospects in auto-driving,blind reading,smart city,unmanned supermarket,production control,security,information extraction and other fields.However,there are still many challenges in extracting text from images of natural scenes and completing the task of text detection.On the one hand,due to the different imaging factors such as shooting distance,illumination,angle,pixel,etc.,the font size and color are often changeable,and the text is skewed.On the other hand,the background of the scene text image is variable,and there are many patterns or textures similar to the text.This also causes the poor performance of the scene text detection algorithm.At present,the task of text detection in natural scenes is challenging.In recent years,in-depth learning-related methods have been frequently used in problem solving in the field of computer vision,and the detection effect is much better than that of traditional methods.This thesis introduces the principles of the existing text detection and recognition technology based on in-depth learning,and then focuses on the analysis of the convolution neural network-based text detection technology for natural scenes,which ultimately completes the recognition and processing of text targets in natural scenes.In this thesis,aiming at the scale problem of target text in natural scenes,the characteristics of existing detection and text correction algorithms are summarized,and the advantages and disadvantages of the key steps are analyzed.Finally,the text detection model and text recognition model are given in combination with the actual situation.The main contents of this thesis are:1.For the existing text detection algorithms based on convolution neural network,such as small-scale incomplete text detection and large text target detection area,an improved YOLOv3-based text detection algorithm is proposed.First,the basic feature extraction network DarkNet-53 is improved to suit the characteristics of natural scene text,and then the detailed features are extracted.Then,combining with multiscale prediction network,this thesis gives a text location prediction method.Finally,the training of the detection algorithm model is completed and a comparative test is carried out.By comparing the benchmark detection methods in relevant literature,the validity of the proposed text detection algorithm model is verified.The experimental results show that the improved algorithm can achieve a wider scale text detection task in natural scenes,and has some advantages over other commonly used benchmark methods.2.A DenseNet based skewed text recognition algorithm is presented for skewed distributed text in natural scenes.Firstly,according to the key point information of text distribution,the text skew information is calculated and the corrected text image is predicted.Then,on the basis of correcting text images,a text recognition network is constructed by using dense convolution instead of traditional convolution,and a text recognition model for natural scenes is trained.Finally,the validity of this text recognition model is verified by comparing the methods of text recognition in related literature.The experimental results show that the improved text recognition model can reduce part of the impact of skewed distribution of text on the recognition results in natural scenes,and the dense convolution network model has better performance for text recognition.