| With the rapid development of network communication,digital media technology and artificial intelligence,a large number of digital image resources have been generated.Presently,analyzing these colorful image resources and extracting valuable information are the research focus in the field of computer vision.Because the text in the natural scene image contains rich semantic information,it can help people communicate and learn better.Therefore,it is of great significance to detect and recognize text in images and videos in natural scenes.This paper focuses on the text detection algorithms in natural scene images.Firstly,the research background of scene text detection and the research status at home and abroad are introduced,then the current popular text detection models are introduced and analyzed,and the evaluation criteria of text detection models are discussed in detail.Based on the text area obtained by the detection algorithm,this paper also analyzes and implements the classic text sequence recognition algorithm.Aiming at the problem of poor accuracy of text detection and recognition algorithms in natural scene images,the main research contents are as follows:(1)Discussing the existed end-to-end text detection model EAST,analyzes the shortcomings of the EAST model,improving the network structure of the EAST detection model,and using the Refine Net structure to fuse features to solve the problem that the EAST model cannot completely detect long text due to the limitation of the receptive field.Secondly,the Loss function Loss of the model was adjusted,the problem of uneven weighting of the data set samples was improved,and the accuracy of text detection was improved.At the same time,the use of Res Net50 as the basic network improves the robustness of the model in complex scenarios.(2)Discussing the related problems of scene text recognition,using improved CRNN recognition algorithm to recognize the text area,using Dense Net model as the convolution layer,and the method focused on the recognition of Chinese text.At the same time,affine transformation is used to adjust the text direction for non-horizontal text in natural scene images,which improves the accuracy of text recognition.(3)Comparing and verifying the effect of improved scene text detection algorithm and text sequence recognition algorithm in different text data sets.Experimental results show that the proposed natural scene text detection andrecognition algorithm can detect the text position efficiently and accurately,and has good robustness to complex scenes.Comparing with other methods of text detection and recognition,the speed and accuracy of the algorithm have achieved good results,which have important theoretical research and use value. |
