Deep Learning Based Scene Text Detection And Recognition

Scene text detection and recognition technology is a hot research field in computer vision.On the one hand,as one of the most important subtask to achieve complete image understanding,the scene text detection and recognition technology have high research value.On the other hand,in terms of application scenarios,high-performance text detection and recognition systems have important practical significance in many fields such as document analysis,smart city or industrial inspection.Firstly,this paper analyzes some representative works of text detection and recognition approaches based on both traditional machine vision technology and Deep Neural Networks(DNN),and then elaborately summarizes their modeling approaches,highlight ideas and core technologies.Secondly,based on the above study,this paper proposes six specific improvement measures for the deficiencies in the existing text detection and recognition methods.In terms of text detection,in view of the lack of modeling of the boundaries of text objects in mainstream methods,this paper proposes a novel text detection approach based on watershed segmentation and designs a complete text detection pipeline based on this proposal.In terms of text recognition,Recurrent Neural Networks(RNN),which are generally used in many mainstream recognition methods,are computationally expensive and difficult to train.In view of these problems,this paper proposes a pure convolutional text recognition method.The main works of this paper are listed as follows:(1)Research on efficient watershed segmentation based scene text detection.Firstly,this paper proposes a novel watershed segmentation based text detection method in view of the shortcomings of the current mainstream text detection algorithms that cannot properly model the boundaries of text regions.Secondly,this paper design six specific data augmentation methods,such as rotation,scaling,stretching,cropping,resampling and color transformation against the objective requirements of the text detection method proposed by the text.Thirdly,aiming at the characteristics of the loss function of the model,a specific Online Hard Example Mining(OHEM)method is designed.(2)Research on efficient fully convolutional network based scene text recognition algorithm.Firstly,for the problems of long gradient propagation path and high computational cost in multi-layer RNNs commonly used for text recognition,a novel text recognition method based on pure CNN is proposed.Secondly,for the deformation problems such as rotation and distortion existing in scene text images,an adaptive image correction based on Spatial Transform Network(STN)is introduced.(3)Comprehensive experiments on the above text detection and recognition methods.This paper performs sufficient experiments on the proposed text detection and recognition methods and optimization strategy on ICDAR2013,ICDAR2015,and TD-500.The experiment proves that compared with the mainstream methods,the computation cost of the text detection method proposed in this paper is relatively reduced by 43.03%,and the detection accuracy is increased by 9.59%.The computational cost of our proposed text recognition method is relatively reduced by 73.47%,and the recognition accuracy is relatively improved by 9.75%.