The Research And Application Of Scene Image Text Extraction Method

Scene text contains important semantic information of scene images.For example,billboards beside streets,traffic signs,signboards of shops all contains scene texts,they can point out related geographical position information.As well as captions and authors on the covers of books,products' names on the packaging can indicate the content of the images.So scene texts will be useful for scene image analysis, browsing and retrieval to extract text information.And it can also be applied to robot vision and blind man vision applications.This paper mainly introduces the related aspects and steps on scene text extraction.The research work is focus on text region localization and binarization in scene images,including description of concepts and algorithms.At first,related work is to extract texts from scanned images,and then more and more people started to research video ocr.Compared to scanned images and video frames,scene images have various resolutions,more complex background,and shooting deformations,which bring scene text extraction work difficulties and challenges.In text localization,we propose a hierarchical block filter and edge feature clustering based method.First,we use the edge detector to find edges appeared in the scene images and making use of block filtering and region filtering to generate text regions,while edge feature clustering will combine text blocks to text item regions. Filtering on different levels can have high recall rate,at the same time,it can decrease the error rate.Edge feature clustering can link semantic related text blocks together. The restoration from scene text deformation can improve binarization results.In text binarization,this paper propose two improved methods:text stroke width feature clustering based method will let stroke width be the main feature to classify text pixels and nontext pixels.Clustering result is good but the number of iterations is big,so it will cause long computation time.While text stroke maker image combination based method will find stroke position approximately,and form final result through combining traditional method binarized result.Its computation time is fast and it also makes use of text stroke feature.Algorithms proposed in this paper are all verified by well designed experiments. The experimental results show that the algorithms proposed in this thesis can achieve high performance both efficiently and effectively.As an application,the whole text information extraction processing has been exploited in the multimedia information retrieval project developed by our lab,and we also hope that these algorithms will be applied to lab's robot vision applications in the feature.