Research On Text Detection And Recognition Method Of Natural Scene Based On Deep Learning

Under the natural scene,people come into contact with every day much more special image,such as advertising,posters,road signs,house number,etc.,in these images,and contains a large number of text messages,people need through to the image in terms of information transmission,can make people communicate more convenient,also can better to understand the world.However,natural scene image has the following features: text language is different,different font size,arranged in a different direction,the interference of background information,the image is often caused by external factors such as light,keep out the resolution is not high,and so on,the traditional OCR techniques can achieve good accuracy at present,but it in view of the image in the text is very neat,backgrounds and text can be separated from the gender is tall,so text information extraction from natural scene images is a very challenging research work.According to the characteristics of natural scene images,deep learning method,which is widely studied at present,is adopted to start from the two aspects of text detection and text recognition.The following is the specific research method:(1)Text detection work is written in the image area to be marked,and in view of the same text text,its size,direction and distance are different,so think about the level of vertical direction than in predictive text direction easier,finally USES is CTPN text detection algorithm,CTPN is a detection algorithm based on CNN is combined with RNN,CNN a convolution is utilized to extract image features,RNN on the characteristics of the character sequence recognition.(2)After finished text detection work,need to identify marked text box,this paper USES the convolutional neural network is used to identify the text DenseNet,compared to other convolution network,the biggest characteristic is intense connection,can alleviate a certain degree of gradient disappeared,also can reduce the quantity,make the calculation faster and identification efficiency is higher.Finally,the CTC algorithm is used to segment and integrate the sequences of single words.Character recognition system based on the natural scene Tensorflow and Keras depth study on the framework,run calculation,according to the process of training and testing,was able to distinguish a good text in natural scene images,compared with other methods,the character recognition is improved,suggesting a CTPN + DenseNet + CTC text recognition in natural scene image has a better effect.