Research On Text Localization Method In Natural Scene Pictures

The study of text localization in natural scene is a difficult problem in the field of text detection.With the development of science and technology,the source of information acquisition is no longer limited to books and magazines.There are abundant information in natural scenes that can be seen everywhere.So how to get the text information in natural scenes quickly and accurately is becoming particularly important.At present,many experts and scholars locate text based on various features.However,the accuracy of text localization is still not ideal due to the influence of natural background,complex background environment and shooting equipment.For the text deformation caused by different viewpoint and applicable to different language texts in natural scene image,this paper proposes a text localization method fusion stroke width variation and maximally stable extremal region.Using the affine invariance of maximally stable extremal region and the applicability of stroke width variation to acquire candidate text area in the image.Then filter the candidate text areas and finally get the good localization effect.It mainly applies the following methods: image grayscale processing,image marginalization,maximum stability,extremal region feature extraction,stroke width transform feature extraction and priori information filtering.This paper mainly includes the following aspects of the work:(1)This paper introduces the background and current situation of research on text localization in natural scenes at present,analyzes the principle and experiment process of main scene text localization method,summarizes the characteristics and advantages and disadvantages of each of the classic method which are seemed as the theoretical basis of the feature selection in text localization method.(2)This paper uses a text localization method that based on the maximum stable extremal regions features.It uses the affine invariance and high robustness characteristics of the maximum stable extremal regions and can solve the problem of text distortion caused by the indeterminate point of view of natural scene image.It firstly processes the image to grayscale,and extracts the maximum stable extremal regions by calculating the area changing ratio of image in the process of threshold change.Then it filters the non-text domain using the prior information and so on to get the final result.(3)This paper proposes a text localization method based on maximum stable extremal regions and stroke width transform feature,and optimizes the acquirement of stroke width.It preserves the affine invariant of maximally stable extremal region features and integrates the character of stroke width.It uses the feature that applies the text of various languages and fonts,and improves the scope of application of the localization method.Firstly,it preprocesses the scene image and then acquires the maximum stable extremal region feature and optimized stroke width transform feature,generates candidate connected domain and filters non-text domain.Finally merge candidate connected domain to generate continuous text area.Experimental results show that this method can solve the problem of text deformation in natural scene images,and it can also be applied to all kinds of languages and forms of text.It can locate the text in natural scenes accurately.In the future,we can further study some special scene environments,such as improving the accuracy of text area acquisition under strong light or dim light.We can focus on reducing the computational cost and time complexity of text localization.