Research On Nature Scene Text Detection And Recognition Algorithms Based On Deep Learning

Text,as one of the most influential inventions of humanity,has played an important role in human life,so far from ancient times.The rich and precise semantic information carried by text is very useful in a wide range of vision-based applications,therefore text detection and recognition in natural scenes has become more and more important.It becomes an active research topic in computer vision and document analysis.Especially in recent years,a surge of research efforts and substantial progresses have been developed in these fields,though there are a variety of challenges(e.g.noise,blur,distortion,occlusion and variation etc.).For this purpose,we study those problems in scene text detection and recognition and obtain the innovative achievements as follows.For scene text detection,we proposed a Multi-Channel and Multi-Resolution Maximally Stable Extremal Regions(MC-MR MSER)based candidate extraction and coarse-to-fine filtering method to detect text in scene images.First,we extract MSERs as text candidates with a proper MC-MR MSER strategy.Then,we design a coarse-to-fine character classifier to discard false-positive candidates,where the coarse filter is based on morphological features as well as stroke width,and the fine filter is well-trained by convolutional neural network(CNN).Finally,horizontal and multi-direction text strings are formed with a graph model on remaining characters.The proposed method is evaluated on ICDAR2013 Robust Reading Competition benchmark database Challenge2 and the practical challenging multi-orientation scene text database(USTB-SV1K)with standard rules.Experimental results show our method is efficient and effective.It achieves F-Score at 83.84% on ICDAR 2013 database and 51.15% on the more challenging USTB-SV1 K database,which are superior over several state-of-the-art text detection methods.In order to solve the problems of text recognition in scene images,we propose a new text recognition method based on context of latent segmentation through transforming the text recognition into the task of sequence labeling based on the development of deep learning technologies.First,the input image is pre-processed to conform the network structure.Then,we leverage CNN to generate the sequential feature of the whole word image.After that,a Bi-directional long short-term memory(Bi-LSTM)recurrent model is developed to robustly predict the generated feature sequences.Finally,we transform the prediction to word with connectionist temporal classification(CTC).We evaluate our algorithm on ICDAR2013 challenge 1,2 and 4.The experimental results show the proposed method has a high recognition performance and speed.Based on the natural scene text detection and recognition algorithms,we proposed an end-to-end scene text detection and recognition system.In addition,for multi-direction text,we extract the direction of the text,and then tilt correction the text with its orientation,which effectively improves the recognition rate.Since the word contains more semantic information than the character,we combine the text recognition algorithm with the result of text location method to improve the accuracy of text localization.