Research Of Key Techniques In Scene Text Detection

With the rapid development of information technology,a large amount of information is disseminated in the form of images and videos.The information society has become a trend of social development.With this,large-scale processing of massive image data has become a problem.With the promotion of artificial intelligence,technologies such as face recognition,intelligent security,and intelligent transportation have brought a lot of convenience to people's lives.I believe that in the near future,the extraction of information in images and videos must have a lot of use.Under such conditions,technologies such as text detection and recognition have rich application fields and broad market prospects.At the same time,they also propose higher technical requirements.Although the traditional optical character recognition technology has been very mature,but the text detection and recognition technology in the natural scene can not meet the growing needs of people.In this paper,the text region detection technology in the image is discussed.The popular target detection algorithm is used to apply in the text region detection field.The main work of this dissertation is as follows:1.The current research status of text area detection in natural scenes is briefly introduced,and the application of maximum stable value region,stroke width transformation and deep learning in this field is emphatically introduced.2 In recent years,the theory of deep learning has been widely used in the field of target detection.Many excellent algorithms have emerged.The detection speed and accuracy have greatly improved.This paper selected the SSD algorithm.First of all,we analyze the characteristics of the text area statistically.At the same time,we cluster the text boxes and adjust the model parameters according to the characteristics of the text area.For the complexity of natural scenes,we have enriched the training dataset by synthesizing some natural scene image data with text regions.We first split the image to get a closed area,and then put the text into the appropriate area.This method can better integrate the text into the image scene.3.There are still some noise disturbances in the detected text area.For better identification,further processing is required.According to the detection result of thetext box,this paper uses the stroke width transformation algorithm to extract the strokes in the text box and filter the noise in the text box.The stroke width conversion algorithm is roughly divided into three steps:the first step is to perform edge detection on the image,the second step is to find the pixel pair in the gradient direction of the edge pixel,and to assign the pixel point between the pixel pairs,the third step is to detect the Components are filtered.Here,we perform two search operations on the positive and negative directions of the pixel gradient,respectively,and use the result of a large number of stroke components in the two search results as the result of the stroke component extraction.This article defines a component filtering method:filter the component by the uniformity of the pixel assignment in the stroke component,the number of other strokes contained in the stroke component bounding box,and the aspect ratio of the stroke component get the final stroke area.Experiments show,this text area detection algorithm has a higher accuracy,and at the same time stroke width transformation in the text area can filter noise to some extent.