Fast Text Detection And Recognition In MOOC Videos

At present,MOOC(Massive Open Online Courses)has attracted a large number of people to participate in these courses,which has aroused the interest of many researchers Most of the current researches focuses on data analysis,such as drop-out rate,pass rate,test score prediction and so on.However,there is few researches on video content,in which text is the main carrier of information.If it can be detected and recognized accurately,it will be very helpful for MOOC video retrieval,video content summary,video classification and other tasksDocument text,that is,the text exists in various documents and books.Because of the simple background and single font,the task of document character recognition has been well solved and applied.Scene text,that is,the text that exists in the image of the natural scene,including signboards,packaging,road signs and so on.Because of the complex factors such as font,scale and background,scene text recognition is a challenging research,and it is much more difficult than document text recognition.At present,many studies also focused on scene text recognition.Similar to the text in MOOC video,it's complexity lying between the above two,but been rarely studied.The method of document character recognition is too simple to meet the accuracy requirement of MOOC text recognition Scene text recognition method is too complex to meet the speed requirement of MOOC text recognition,because it relies on intensive computationsWe have carried out a series of research on the process of text detection and recog-nition in MOOC,which are stated as follows(1)A fast text detection algorithm is proposed in this paper.Firstly,we propose a candidate character detection method based on pruning maximally stable extremal region(MSER).Unlike existing algorithms in algorithm flow,we replace a traditional classier with a row-level character clustering algorithm to classify characters,Because it can reduce the computational complexity of the whole algorithm process(2)In order to recognize characters more accurately,we use the method of deep learning to construct recognition network.The network consists of CNN and RNN Convolutional neural network(CNN)is used to extract features from images.Recursive neural network(RNN)is used to solve the context relationship between words and capture temporal features.This method can achieve high accuracy,but it will suffer from the problem of slower speed.(3)In order to solve the above problems,this paper proposes three basic primitive operations to build a network from the design of a compact and efficient neural network.Based on these primitive operations,the above-mentioned text recognition network is reconstructed.Experiments show that our recognition network can achieve faster recog-nition speed without losing the accuracy of text recognition.Based on the above research,we propose a fast text detection and recognition algo-rithm for MOOC videos,which also lays a foundation for the in-depth study of MOOC video content.