Scene Text Localization And Recognition Algorithm Based On Convolutional Neural Network

With the rapid development of information technology,images have become the most convenient and informative cognitive carrier for human beings.And use computer vision to complete the positioning and recognition of image text information in natural scenes.It can not only further improve the efficiency of information retrieval and acquisition,but also lay the foundation for the future of artificial intelligence.In recent years,convolutional neural networks have continued to develop in depth,and have significant advantages over traditional document text positioning and recognition technologies.Scene text localization and recognition algorithms based on convolutional neural networks have become a hot research direction in text processing.This paper has done research on the excellent text positioning algorithms in recent years.Aiming at the characteristics of natural scene text,an accurate and efficient scene text detector(EAST)with outstanding text positioning effect is selected as the basic algorithm for improvement.Therefore,the main researches content of this paper are as follows:1.Aiming at the problem that the EAST text positioning algorithm is in sensitive to text positioning of variable length text and complex background text.This paper studies the embedding of channel attention module and spatial attention block in the convolutional layer of the feature extraction branch of the EAST network.It improves the sensitivity of the model to the text area and improves the text extraction effect.Experiments show that in complex text regions and scene text images with large scale changes,the improved method effectively improves the recall and accuracy of text detection compared with the original algorithm.2.In order to further improve the incomplete detection of text areas with large scale changes and the missing detection of small text instances,the previously improved channel attention module and spatial attention module are retained,and the two are changed from a series processing structure to a parallel type.Structure,adding two 3×3 cavity convolutions at the same time to expand the sensing field and reduce the loss of underlying features in the feature extraction process.Experimental data proves that the improved algorithm further improves the accuracy and recall rate of text detection with large scale changes,and also improves the detection rate of small texts.3.For the text recognition and verification work after positioning,this article implements a sequence-based text recognition method.The network structure is a codec structure with an attention mechanism added.The attention mechanism is added to ensure the relevance of the context and effectively improve the shortcomings of poor recognition of long text sequences.Experimental results show that the accuracy of the model in the recognition of scene text can meet the needs of text recognition.