Research On Text Detection And Recognition Under The Natural Scene

How to make computers understand images accurately has always been a challenging topic in computer vision.Natural scene images usually contain non-text elements and text elements.The text elements contain rich semantic information and recognize the text elements can help the computer to understand images more accurately.Word recognition technology has also been widely used in human-computer interaction,license plate recognition,driverless,automation and other fields.Optical character recognition technology(OCR)has achieved remarkable results in document scanning,handwriting recognition,etc.after decades of development since the interview.However,the traditional OCR technology can not be applied to the text recognition in natural scenes because the text recognition in the natural scenes is faced with difficulties such as complex backgrounds,fickle variations and low resolution.The traditional OCR technology almost can not work in these cases.With the development of related technologies and the needs of related fields,the breakthrough and development of text recognition technology in natural scenes have academic and practical value.This paper has a research on the text detection and recognition in natural scenes which based on a large number of references and relevant literature,summarizes the current research status in relevant fields both at home and abroad,combined with related technologies such as Maximally Stable Extremal Regions(Maximally Stable Extremal Regions,MSER)text detection algorithm,Convolutional Neural Network(Convolutional Neural Networks,CNN)and Recurrent Neural Network(Recurrent Neural Networks,RNN).This paper mainly consists of two parts,summarized as follows:The first part is about the text detection algorithm in the natural scene.This paper introduced the MSER text detection algorithm in detail firstly.It is found thatthe MSER text detection algorithm has a better detection rate in the case of high contrast between text and background,however,when the contrast between the text area and the background is low,the effect of text detection is not satisfactory through experiments.By analyzing the principle of selecting the extreme value region deeply,this paper improves the traditional MSER text detection algorithm and relaxes the condition that the MSER selects the extreme value region so that it can detect the text region with low contrast between the text and the background.The second part is the study of text recognition algorithm in natural scenes.This part introduces the deep learning framework in detail firstly,and then builds a text recognition model composed of CNN and RNN.This model extracts the static features by using CNN,extracts the context features by two reverse short and long-term memory networks(LSTM),and combines the two features into a feature vector sequence as the input of the decoder,which passes through the decoder Intentional mechanisms are introduced to get the position of each character in the image.In this paper,ICDAR-2013 dataset is used to train the network.Finally,experiments on the test dataset show the effect of the model text recognition.