Application Of Convolutional Deep Belief Network In Text Recognition

Lots of text information was contained in natural scene image,and the identification of these information was of great significance in the retrieval management of massive information such as-video,image and picture.Because of these problems such as complex image background,low resolution,variety of font and random distribution in text recognition of natural scene images,it still was a challenging task.Therefore,the study about text information recognition in natural scene was done in this paper.And the Main contents were as follows:1.The method of natural scene text detection which combines the Maximally Stable Extremal Regions algorithm with the Convolutional Deep Belief Network was proposed.First,Maximally Stable Extremal Regions algorithm was used to locate the text of the natural scene image,and then the preprocessed candidate text area was sent into the Convolutional Deep Belief Network for training.More hidden features was learned from the training data,and inputted into the Support Vector Machine to discriminate,verify the candidate text area.Finally,a large number of non text regions were filtered out,and text detection results were got.2.Using the Convolutional Deep Belief Network to recognize the text area after detection was proporsed.First,the extracted text area was input into the Convolutional Deep Belief Network for feature extraction,and then the extracted feature was used as the input of the Softmax character classification recognizer to output the corresponding characters or words.At last,the recognition results were processed to get the final output.3.The text recognition method was tested on the ICDAR2011 dataset and the SVT dataset,and the influence of the iteration number,the number of hidden layers,the introduction of the random noise and the different pooling methods on the performance of the Convolutional Deep Belief Network was analyzed.The results of these experiments showed that the natural scene text recognition method based on Convolutional Deep Belief Network could achieved good results on ICDAR2011 and SVT datasets.On ICDAR2011 dataset the accuracy rate of text detection was 90.84%.The accuracy rate of text recognition was 92.05%.The accuracy rate of the SVT dataset was65.72%.And the accuracy rate of text recognition is 76.16%.