Studies Of Scene Text Detection And Recognition Based On Deep Learning

In our daily life,text detection and recognition has a very wide range of applications,including image retrieval,image control,ticket recognition and many other scenes.As a result,it has attracted extensive attention in the academic circle,and many experts and scholars have conducted further research on it.In natural scenes,texts usually have different types,scales and distributions,and lack a structured layout.In addition,the complexity of the background of scenes may affects the accuracy of the results of detection and recognition.For examples,low resolution,noise interference,foreign body shielding,perspective transformation,etc.,make text detection and recognition in the scene a challenging task.Based on the existing algorithms,this paper studies and improves some algorithms.In order to cope with poor detection effect in complex scenes,two new text detection algorithms,based on convolutional neural network,are put forward in two different application scenarios.Meanwhile,a new algorithm of text recognition based on corrective network is proposed so as to solve the problem of low efficiency of text recognition in complex scenes.All the three algorithms above have been tested on the standard data set and obtained satisfactory results.The main research contents of this paper are as follows:1.First,text detection algorithm based on multi-scale feature fusion and pixel point link is proposed.This algorithm mainly solves the problem of low efficiency of anchored point algorithm.Compared with the algorithm based on anchor point class,the algorithm is based on instance segmentation,with a simple the process and a higher detection accuracy.Since the algorithm does not need to set anchor points,it saves computation and improves detection efficiency.The algorithm has achieved good results on standard data sets ICDAR2013 and ICDAR2015.2.Second,edge pixel based on tilting and long text detection algorithm is proposed.This algorithm mainly solves the problem of poor detection effect of tilting and long text.In this algorithm,the rotation candidate box with angle is adopted and the edge pixel is introduced to assist the detection.The final candidate box is generated by the edge pixel,rotation angle and the boundary distance.Good results are obtained on standard data sets with more slanted long text,which fully demonstrates the excellence of the algorithm.3.Scene text recognition algorithm based on corrective network is proposed.This algorithm mainly solves the problem of large scale difference and uneven distribution.In this paper,a correction network is used to correct the picture to a more recognizable position,which improves the efficiency of the network.The sequence recognition part adopts the bidirectional short-time memory network,which is sensitive to the context information,so it can be used to extract the information between characters for modeling.In addition,this algorithm also introduces an additional decoding layer in the decoding layer,which is used to add some additional supervisory information to assist the recognition,thus improving the speed of recognition and the robustness of the algorithm to some extent.At the same time,the algorithm performs satisfactorily on data sets such as ?T5 K.