(复杂背景下文字的定位和识别-国际顶级会议ICCV上的)End-to-End Scene Text Recognition-wang_iccv2011.pdfVIP

End-to-End Scene Text Recognition Kai Wang, Boris Babenko and Serge Belongie Department of Computer Science and Engineering University of California, San Diego kaw006,bbabenko,sjb@ Abstract This paper focuses on the problem of word detection and recognition in natural images. The problem is significantly more challenging than reading text in scanned documents, and has only recently gained attention from the computer vision community. Sub-components of the problem, such as text detection and cropped image word recognition, have been studied in isolation [7, 4, 20]. However, what is un- clear is how these recent approaches contribute to solving the end-to-end problem of word recognition. We fill this gap by constructing and evaluating two sys- tems. The first, representing the de facto state-of-the-art, is a two stage pipeline consisting of text detection followed Figure 1. The problem we address in this paper is that of word by a leading OCR engine. The second is a system rooted detection and recognition. Input consists of an image and a list of in generic object recognition, an extension of our previous words (e.g., in the above example the list contains around 50 total work in [20]. We show that the latter approach achieves su- words, and include ‘TRIPLE’and ‘DOOR’). The output is a set perior performance. While scene text recognition has gen- of bounding boxes labeled with words. erally been treated with highly domain-specific methods, our results demonstrate the suitability of applying generic Despite its apparent usefulness, the scene text problem computer vision meth