Text Recognition In Natural Scene Based On Deep Learning

The natural scene text recognition technology is one of the research hotspots in the field of target recognition at present.It is widely used in unmanned driving,blind assistant system,traffic security and other fields.Although the text recognition has been greatly developed,the background of text images in natural scenes is more complex,and the font size is variable,and the text direction is tilted,which makes the current research methods have poor recognition effect.In recent years,deep learning has been widely used in text recognition and has achieved some results.Therefore,the research on natural scene text recognition based on deep learning has high value and significance.Firstly,this thesis elaborates the research status of related fields at home and abroad as well as the difficulties of the current natural scene text recognition technology,and deeply studies the related theoretical knowledge.On this basis,the text detection and text recognition algorithms based on deep learning are analyzed,and the text recognition scheme based on deep learning is designed.Then,according to the problem that the text size of the natural scene is variable and the text detection is not good due to the tilt of the text direction,a text detection algorithm based on RFPN-RCNN is designed.The algorithm extracts multi-scale features of text by recurrent pyramid network,generates multiple proposed boxes by the multi-directional window extraction network SRPN,and oblique proposed boxes are obtained by rotating the proposed boxes module,and the final text boxes are generated by filtering the oblique boxes.The experimental results show that the text detection algorithm designed in this thesis can effectively detect the text in the natural scene.Finally,aiming at the problem of poor performance of text recognition algorithm for the tilted text recognition,this thesis proposes a text recognition algorithm based on TPC-EDN.The algorithm uses the self-learning characteristic of CNN to modify the tilted text into horizontal text which is easy to recognize.The text content can be accurately recognized through the EDN model designed in this thesis.The EN module applies dense connection network and BLSTM to extract the space and sequence characteristics of text effectively,and generates coding vectors.The DN module converts coding vectors into output sequences through attention mechanism and LSTM.The experimental results show that the text recognition model designed in this thesis improves the recognition rate of natural scenes.