Research On Scene Text Localization

Scene text localization is a process of locating text areas in natural scene pictures,and is a branch of scene text recognition.It is important for extracting text information in natural pictures.It is widely used in pattern recognition,machine vision and other fields.As the rapid development of the Internet,the scene text localization is also very important for the correct understanding of massive picture information,building content-based image retrieval.However,the background of the natural picture is complicated,the text is diverse,and it is easy to be affected by factors such as light,shadow and low resolution,which brings a lot of challenges to the localization of text.In this thesis,we study a large number of domestic and foreign literatures,and propose a method to combine texture image features and connected domains for scene text localization.Firstly,the image is processed by extracting the Maximally Stable Extrernal Regions,and the regions which are obviously inconsistent with the characters rule are selected by some rules.Meanwhile,the image is transformed into the stroke width figure.The two methods are filtered to merge the remaining regions to get the candidate character concatenation domain.Then,the candidate character connected domain is scaled to uniform size.The gradient histogram feature and local binary pattern are extracted by using sliding window,and the texture feature vectors are inputted into the trained SVM.Then the SVM distinguish the connected domain.The classifier decides that the non-word connected domain is removed,leaving only the connected domain of the classifier as the text.Finally,according to the rule of width,height,area and color of the character area,the residual character connected domain is expanded and merged into the connected line of the text,which is the output result of the algorithm.In this paper,the algorithm is evaluated according to ICDAR evaluation standard,and the dataset is ICDAR2011 and 2015.And compared with other algorithms in the same dataset,the advantages and disadvantages of the algorithm are analyzed.Experimental results show that this algorithm can locate the text in the scene picture,and the recall rate is high.