Research And Application Of Text Detection And Recognition Technology Based On DBSCAN

Text detection and recognition have been widely used in image retrieval,real-time translation,information filtering and other fields.Nowadays,it has become a popular research topic in computer vision,artificial intelligence and other fields.Text detection and recognition technology is mainly composed of text detection and text recognition.Text detection is the precondition for text recognition,which has a crucial influence on the recognition accuracy.There is no doubt that text detection is a mature technology when the text region is horizontal or vertical.However,text detection is still a challenge in text region of arbitrary shapes,especially in the curved text region.For text recognition,the accuracy of traditional Optical Character Recognition(OCR)for document image has reached nearly 100%,but when it is applied to the scene text in images,because of background interference,image quality,the recognition efficiency will be greatly reduced.These factors have brought great difficulties to the text detection and recognition of images.This dissertation studies the techniques of text detection and recognition,a text detection algorithm based on DBSCAN clustering algorithm is proposed for arbitrary shape of text region.And then filter and correct the curved text region,which can improve the recognition accuracy of the curved text region greatly.The main work includes the following aspects:(1)This dissertation proposes a text detection algorithm based on DBSCAN clustering algorithm for arbitrary shape of text region.First of all,preprocessing the image includes image graying,histogram equalization,image smoothing,image binarization.Then,the position information of the pixel in the image is converted to the data point of the two-dimensional.Finally,according to the features of text in the image,cluster the points by using DBSCAN,and the text candidate regions is obtained according to the clustering result.This algorithm can filter non-text region effectively,and it is an innovative idea for detecting texts with arbitrary shapes.(2)Aiming at the nested text candidate regions,a method to filter the nested regions by area comparison is proposed.Because of the background interference in curved text region,this dissertation designs a filtering method based on optimized convex hull algorithm.On the basis of optimized convex hull algorithm,this dissertation optimizes the parameters value method of the Alpha Shape algorithm and used for further filtering.Finally,the background interference in curved text region is reduced effectively.(3)In order to improve the accuracy rate of curved text,this dissertation designs a correction method for the curved text region based on affine transformation and log-polar transform.This method consists of three steps: characters segmentation,coarse processing and fine processing.The first step is to split characters by projection.Then the second is to rotate text region so that it can be placed horizontally by affine transformation.The last step is to correct the curved part in text region by using log-polar transform.After processing,the curved text region becomes horizontal.(4)Finally,AI open platform are used to recognize the text region,and a contrast experiment was conducted.The experimental results show that the algorithm which is proposed by this dissertation not only improves the recognition accuracy of the curved text region,but also has great theoretical significance and practical value.