Fast Classification Of Web Document Images

With the rapid development of Internet,smart phones and communication technology,multimedia data such as texts,images,videos and audios on the Internet increases rapidly,which brings rich information and great convenience to our life.On the other hand,it is becoming more and more difficult to exploit the information embedded in the heterogeneous data.Specifically,considering the increasing proportion of the visual data in the network data,we can also find that the number of document images grows rapidly in recent years.The recognition and understanding of document image content is of great significance for the effective use of network information.In this paper,we focus on the fast classification technique for web images,which is studied and applied to classify the web images into five categories,namely natural scene text / non-text images,born-digital images,camera-captured paper documents and scanned paper documents.To end this,we build two classification systems,the first one is type-based image classification module and the second is a content-based module.Specifically,the type-based classification system consists of two stages: the first stage extracts global features reflecting distributions of color and saturation and uses a support vector machine(SVM)classifier for classification.The images assigned low confidence by the first-stage classifier are processed by the second stage,which extracts local texture features represented in the Bag-of-Words framework and uses another SVM classifier for final classification.The contentbased module is used to distinguish text and non-text images in natural scene images.It first detects and localizes the Fastext corners in the natural scene image and then filters them using the combination of local and global non-maximums suppression.Finally,a flood fill operation is performed based on the candidate text points to generate candidate text connected components(CCs).The geometry features and local texture features are extracted based on those CCs and fed into a text/non-textregion classifier for final decision.Our experimental results on a large image database demonstrated the effectiveness of the two proposed classification methods.Our works have two major contributions.First,we focus on the fast classification technique for web images,design effective and efficient features and build a real time classification system.Second,we build a large image database called NLPR_Web4,which contains over 40,000 Web images.This database can be shared for academy research.