Research On Text Detection And Recognition Methods In Natural Scene Images

Text contains key high-level semantic information in natural scene images.Text recognizing can help computer understand and analyze images more accurately.The research of text detection and text recognition in natural scenes has important theoretical significance and practical value for intelligent transportation,visual recognition,image translation and other fields.There are many difficulties in natural scene images,such as complex background,low resolution of image blur,variable form of single character and so on.In this paper,two specific cases of natural scene are studied: simple scene disturbed by complex background and fuzzy scene with low image resolution.Researches include text detection,character segmentation and character recognition.Our specific work is as follows:(1)A text detection method based on stroke width and pruning algorithm is proposed to solve the problem of complex background in simple scene text image.Text similarity of each candidate area can be obtained by the designed stroke width feature extraction algorithm.Fusion pruning algorithm is used to filter the background area.Target detection image has high accuracy and removes complex background interference regions better.The simple scene text detection is realized based on connected components method.Experimental results show that detection accuracy of this method is more than 0.78,and the effect of filtering background is good.(2)Aiming at the problem that resolution of blurred scene text image is difficult to detect,a text detection method combining the guided filtering algorithm and edge detection is proposed.Dark channel theory and improved mean-based guided filtering algorithm are used to perform de-blurring image enhancement processing which can highlight the edge information of text areas and enhance image contrast.Using image enhancement algorithm to improve the clarity of the fuzzy scene image.Combining the edge information to complete text detection in fuzzy scene image.The running time of this method is fast.The accuracy of detection is not less than 0.85.And the effect of fuzzy scene text detection is very good.(3)Targeting at the problem that form of single character in natural scene is changeable,and character segment and recognition is difficult.The row and column segmentation of single character are carried out for text detection image of simple scene and fuzzy scene respectively.Segmentation of adhesion character in fuzzy scene is completed by improved vertical projection method.The single character is input to the classic CNN(Lenet-5 network) to finish single character recognition.Using the spatial features between characters effectively solves the problem of single character segmentation in natural scenes.And realizing the text content recognition through CNN.The test recognition rate of the text recognition network framework of this paper is more than 0.90,and the character recognition performance is good.In summary,this paper studies text detection in simple scenes,text detection in fuzzy scenes,and character recognition.A number of techniques such as stroke width,connected domain,guided filtering,edge detection,and convolutional neural network are used in research of this topic.The experimental results show that natural scene text detection method proposed in this paper has good detection effect,and text recognition network framework used has good test performance.