Text Localization In Natural Scene Based On MSER And Convolutional Neural Network

The text in natural scene image contains rich and accurate high-level semantic information,which can explain and supplement the natural scene effectively.For this reason,the text recognition in natural scene is a new research hotspot in the field of computer vision,and has been widely used in language translation,unmanned vehicle navigation,unmanned aerial navigation,industrial automation,image information mining,image search,data and other fields.Text recognition in natural scene refers that the computer,smart phones can identify the text like people.The text information extraction in natural scene is confronted with many difficulties,such as complex background,variable font,noise,fuzzy and other factors.In recent years,the text recognition research in natural scenes has made some Results,but there still exits a big gap with the practical application of the requirements.Text recognition in natural scene mainly includes two sub-problems: text detection and character recognition,in which text detection is more concerned.Text detection refers to text information confirming in the image,and then locate and split the text.In this paper,we make research on text detection as below:1.We propose an algorithm based on improved maximally stable extremal regions(MSER)for extracting candidate regions.MSER region detection operator has very good robustness,can extract low-quality character area.But MSER can only operate in the gray channel,and sensitive to light.When the image is dark,there will be character leakage.In this paper,we improve the MSER operator.Firstly we do the light compensation in the R,G,B channels,and then extract MSER area in three channels.Finally the region are merged in three channels as the final character candidate area.Experiments show that the improved algorithm can detect more character areas and improve the recall rate of text.2.We propose a non-character region filtering algorithm based on convolution neural network.There are many elements in the natural scene image,such as doors and windows,railings,leaf meshes,lampposts,etc.,so the character extraction will detect many non-character candidate areas,which is one of most important challenges based on MSER.In order to detect the text area in the natural scene image accurately and quickly,this paper combines the three steps to filter the non-character candidate region:(1)filtering the non-character candidate region according to the basic character of the character;(2)filtering the non-character candidate region based on the character stroke width;(3)filtering the remaining non-text area based on the convolution neural network.Experiments show that this method can locate the text area in the natural scene image accurately,and improve the accuracy of text detection,while ensure the detection speed.