Researches On Text Localization Based On Multiscale Geometric Analysis

As an effective medium to convey information,text has always been focused by people.The key of character recognition technology is how to get effective text region,namely text location technology.In the field of digital image processing,multi-scale Geometric Analysis is an important tool.In this paper,we discuss methods of text localization in images with complex background based on NSCT and Discrete Shearlet Transform,respectively.Specific content is as follows:Firstly,this paper introduces the background knowledge of text location and explains the development history,the application fields and the difficulties in detail.At the same time,it introduces the current situation of the research,then classifies and compares the text location methods in detail.Finally,we describe unusual characters in images and their corresponding solutions.Secondly,the improved text location method based on NSCT is proposed.The method is divided into three parts.First of all,the NSCT is used to perform multi-scale and multi-directional decomposition on the training images,and we perform text region features training.Then the test image is classified into text region and non-text region,and the non-text region is discarded directly.At last,the method performs edge detection to recognize the text-region edges in high frequency sub-bands and morphological operations to connect the scattered edges together.With the help of the logical AND operator,we get the common text regions of edge information images at the same scale.Next,voting-decision is applied among different scales to get the more accurate candidate text regions.The final text regions are got by heuristic knowledge.Thirdly,because the Discrete Shearlet Transform algorithm is simple and flexible,the new improved text location method based on Discrete Shearlet Transform is proposed.This method is similar to the text localization method based on NSCT.First of all,the Discrete Shearlet Transform is used to perform multi-scale and multi-directional decomposition on the training images,and we perform text region features training.Then the region classification is used to discard the non-text region information of test image.At last,the high frequency sub-bands are processed by dynamic thresholding to get more accurate text regions.Next,we get the final text regions by using morphological operations,the logical AND operator,voting-decision and heuristic knowledge.Experimental results show that the proposed text localization methods are effective.Compared with the method based on NSCT,the method based on Discrete Shearlet Transform has the advantages of lower complexity,faster computation and better performance.