Natural Scene Text Localization And Recognition Based On Neural Network

There are many ways to perceive and interpret the world,such as sound,pictures,text,video,et al..Image as the most common way to cover profound language and semantic infor-mation,therefore,the application and research of image correlation technology have become popular,and the text localization of images is also the main field of research.There are many difficulties in text research in scene images,such as complex background,directional text,multi-languages,different brightness and so on.That is why the traditional text localization and recognition methods are difficult to apply on scene text images.This paper explores and research on the use of the traditional method and advanced artifi-cial neural networks to identify and locate scene texts.This article provides a comparative introduction to popular research data sets.A traditional algorithm called MSER feature constraint is proposed,which is used for text localization.And the algorithm is optimized by combining color space to achieved better text localization effect.Analyzed the neural network of scene text localization algorithm from CNN to Mask-RCNN and point out the drawback of each algorithm.Using the traditional scene text localization method of MSER to compare with the algorithm of neural network based Mask-RCNN,it is found that the neural network based localization method has better performance under complex background conditions.The advantage of the traditional method is fast-speed,but it has poor accuracy.The neural network method is highly accurate while the training process is time-consuming.The CRNN network consisted of the convolution,recurrent and transcription layers.Through pre-training of localization network and recognition network separately,and com-bine them to pre-training finally the end-to-end text recognition algorithms based on CRNN are implemented.In the experiment,RNN+CTC network can get the final recognition result.Finally,this paper proposed the WED algorithm combine brand name classification to correct the dataset labels,which fundamentally improves the authenticity of the dataset.