WO2009022290A2 - Procédé d'extraction de texte défilant à partir d'une séquence d'images animées - Google Patents

Procédé d'extraction de texte défilant à partir d'une séquence d'images animées Download PDF

Info

Publication number
WO2009022290A2
WO2009022290A2 PCT/IB2008/053219 IB2008053219W WO2009022290A2 WO 2009022290 A2 WO2009022290 A2 WO 2009022290A2 IB 2008053219 W IB2008053219 W IB 2008053219W WO 2009022290 A2 WO2009022290 A2 WO 2009022290A2
Authority
WO
WIPO (PCT)
Prior art keywords
image
text
images
region
sequence
Prior art date
Application number
PCT/IB2008/053219
Other languages
English (en)
Other versions
WO2009022290A3 (fr
Inventor
Jan A. D. Nesvadba
Bart Kroon
Pedro Fonseca
Original Assignee
Koninklijke Philips Electronics N.V.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Koninklijke Philips Electronics N.V. filed Critical Koninklijke Philips Electronics N.V.
Publication of WO2009022290A2 publication Critical patent/WO2009022290A2/fr
Publication of WO2009022290A3 publication Critical patent/WO2009022290A3/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/60—Type of objects
    • G06V20/62—Text, e.g. of license plates, overlay texts or captions on TV images
    • G06V20/635—Overlay text, e.g. embedded captions in a TV program
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
    • G06V30/10—Character recognition

Definitions

  • the invention relates to a method of extracting scrolling text from a sequence of moving images, including: obtaining video input data representative of a sequence of image frames representing the sequence of moving images, detecting at least one region containing text within the images of the sequence, and, for each of at least one of the at least one detected region: estimating a displacement of the region from image to image.
  • the invention also relates to a use of such a method.
  • the invention also relates to a system for extracting scrolling text from a sequence of moving images, including an interface for obtaining video input data representative of a sequence of image frames representing the sequence of moving images, the system being configured to detect at least one region containing text within the images of the sequence, and, for each of at least one of the at least one detected region: to estimate a displacement of the region from image to image.
  • the invention also relates to a computer programme.
  • US 6,185,329 discloses a text detection method for DCT compressed images.
  • the method starts with the calculation of the horizontal text energy block, where the horizontal text energy is calculated for each DCT block by summing up the absolute value of the DCT coefficients. Then, in a threshold detection block, a predetermined threshold for the horizontal text energy is applied to obtain blocks of high horizontal variation. These blocks are designated as potential text blocks.
  • the potential text blocks are then subject to a morphological operation in a remove noise and merge block to remove isolated noisy blocks and merge disconnected text blocks.
  • the area is provided to an applied temporal adjustment block, where temporal adjustments are applied to consecutive frames which contain text. The displacement of text in two frames is established and candidate frames are eliminated where there are no corresponding areas in the frames before and after.
  • the output is then output as located text.
  • a problem of the known method is that single-image text extraction is relatively inaccurate, in particular where it is applied to compressed images. The inaccuracy can be improved only imperfectly by filtering frames or making up missing text areas. This implies that a relatively powerful Optical Character Recognition (OCR) system will be required to process the located text.
  • OCR Optical Character Recognition
  • This object is achieved by the method according to the invention, which further includes, for each of at least one of the detected regions: registering the images in accordance with the estimated displacement such that the region is generally static from one registered image to the next, and causing at least parts of image frames representing multiple registered images to be used as input in a method of segmenting text from background, based on multiple images so as to generate an image of the text filtered from these.
  • Segmenting text from background, based on multiple images so as to provide a filtered image of the text as output leads to relatively accurate text extraction, because the amount of information on which the filtered image is based is larger than that employed in a single-image text extraction method.
  • a text region is segmented by analysing multiple images in which the text region is potentially present. Due to the step of registering the image frames in accordance with the estimated displacement, such that the region containing text is generally static from one registered image frame to the next, the method is carried out on the basis of image frames representing a sequence of images containing a relatively clearly defined text set against a blurred, because moving, background. As a result, the text extraction is relatively accurate.
  • multiple registered images are blended to obtain at least a partial composite image, such that pixels in at least parts of the composite image have values based on values of the corresponding pixels in the multiple registered images.
  • An effect is that text is made more prominent. In particular, edges become more sharply defined.
  • the effect is prominent where the video input data representative of a sequence of image frames comprises, or is based on, compressed video images, because high- spatial frequency information will generally have been lost in the individual images or image frames.
  • a variant includes applying to the at least partial composite image a colour segmentation algorithm for identifying connected regions similar in at least one of texture and colour, and applying a segmentation method to only those of the connected regions fulfilling at least one certain criterion.
  • Efficiency gains are achieved by examining only those of the connected regions meeting the certain criterion or certain criteria. In particular, non-textured regions of the images can be discarded.
  • the image of the text is filtered from a region corresponding to at least part of a certain one of the connected regions upon determining that the region is dimensioned such as to fulfil at least one criterion.
  • the image of the text is filtered from a region corresponding to at least part of a certain one of the connected regions, identified by calculating a complexity of the region and determining if the complexity of the region exceeds the complexity of at least one other region by a pre-determined ratio.
  • An embodiment of the method includes, for each of a plurality of pixels, computing a variance in value across a number of the registered images, and excluding pixels having a variance above a certain threshold from the generated image of the text.
  • An effect is to take account of the fact that scrolling text is static in the registered images, but that parts that are not moving in a concerted manner in the original sequence of moving images will be moving in the registered images. Thus, in the registered images, pixels across which such parts move will vary considerably.
  • a variant includes identifying those of the pixels having a variance below a certain threshold and a brightness below a certain level in the number of registered images, and applying a morphological operation for closing contours defined at least partially by the identified pixels.
  • An effect is to identify edges of characters. The characters are static over the number of registered images. Character edges are known generally to be quite dark relative to the body of characters.
  • a variant includes identifying those of the pixels having a variance below a certain threshold and a brightness above a certain level in the number of registered images, and excluding from the generated image of the text those of the identified pixels lying outside any of the closed contours.
  • An embodiment of the method includes providing a binary image of the text filtered from the multiple images as output.
  • An effect is to provide output suitable for use by relatively uncomplicated Optical Character Recognition systems.
  • An embodiment includes, prior to registering the images, applying at least one algorithm for enhancing the resolution of the images.
  • the displacement of the region is estimated by means of at least one of: image correlation; video motion vectors; and feature point tracking.
  • Image correlation is a relatively reliable method, involving moderate computational expense. It yields relatively good results when the background is simple and there are no large objects moving in a direction different from that in which the scrolling text is moving. Image correlation also offers the potential for making the method faster by using only selected lines of images in calculations.
  • the video motion vectors technique is relatively fast, especially when using motion vectors parsed from a compressed video stream. It is relatively insensitive to the presence of large objects moving in different directions from the direction in which the scrolling text is moving.
  • Feature point tracking is quite appropriate to estimating displacement of a region containing text, because text is relatively finely structured. It therefore provides many feature points. They are easy to track because the text remains generally invariant as it moves across the image area.
  • the displacement of the region is estimated by obtaining multiple estimates of the displacement of the region, and deriving a single value based on an analysis of the multiple estimates.
  • the multiple estimates may be based on different pairs of images or image frames. Alternatively or additionally, different estimation techniques may be applied. Spurious estimates are avoided, e.g. by averaging or mere discarding of outlying values.
  • the method according to the invention is used in a method of indexing video files for use by a search engine.
  • Scrolling text has a high semantic value. Extracting this text from moving images, e.g. video files posted on the Internet, and subsequent analysis using a fast optical character recognition system and a set of rules allows for meaningful metadata to be obtained for indexing the video files.
  • the system for extracting scrolling text from a sequence of moving images is configured, for each of at least one of the detected regions: to estimate a displacement of the region from image to image, to register the images in accordance with the estimated displacement such that the region is generally static from one registered image to the next, and to cause at least parts of image frames representing multiple registered images to be used as input in a method of segmenting text from background, based on multiple images so as to generate an image of the text filtered from these.
  • the system is configured to carry out a method according to the invention.
  • a computer programme including a set of instructions capable, when incorporated in a machine-readable medium, of causing a system having information processing capabilities to perform a method according to the invention.
  • Fig. 1 is a schematic diagram illustrating a use of a method of extracting scrolling text from a sequence of moving images
  • Fig. 2 is a schematic diagram showing three successive image frames, each representing an image
  • Fig. 3 is a flow chart of a method of analysing the images
  • Fig. 4 is a diagram illustrating a first method of estimating a displacement of a region containing text from image to image
  • Fig. 5 is a diagram illustrating a second method of estimating a displacement of a region containing text from image to image
  • Fig. 6 is a diagram illustrating a third method of estimating a displacement of a region containing text from image to image.
  • Fig. 7 is a diagram illustrating a fourth method of estimating a displacement of a region containing text from image to image.
  • a server 1 is connected to a network 2, e.g. the Internet, and arranged to store video files comprising data representative of a sequence of image frames representing moving images.
  • a computer 3, e.g. a personal computer or a workstation, is connected to the network 2. It comprises amongst others a network adapter 4, central processing unit 5, main memory 6 and data storage device 7. Instructions in software stored in the data storage device 7 enable the computer 3 to execute an application for crawling the network 2 for video files to be indexed for use by a search engine. Indexing involves obtaining metadata for characterising the video files. To this end, an application running on the computer 3 extracts text from the sequences of moving images, in order to provide a binary image of only the extracted text to optical character recognition (OCR) software.
  • OCR optical character recognition
  • the OCR software returns data encoding the text in a machine-readable format, e.g. ASCII or UNICODE, for use by an application for establishing the index.
  • the data may be used for retrieval, summarisation or similar applications.
  • the video files are provided in association with a file including data in a machine-readable format annotating the associated video files.
  • the output data returned by the OCR software is used by a further application to obtain data for comparison with the data included in the files provided in association with the video files. In this way, either the results of the text extraction may be verified or, in case the files provided in association with the video files do not originate from a trusted party, the data annotating the video files can be verified.
  • three frames 8,9,10 represent three successive images.
  • the image frames 8,9,10 each represent an image.
  • two interlaced image frames represent one image, so that each represents a half- image.
  • a first region 11 contains text and moves across the area of display from one image to the next.
  • a second region 12 also contains text, but this text is static.
  • Scrolling text is frequently used for credits at the beginning or at the end of interesting movies, popular soaps or TV series. It contains information about title, actors, director and so on. The text can scroll up or down, left or right. The information has high semantic value and is therefore very discriminating when retrieving and browsing it.
  • Another use for scrolling text lies in the domain of news tickers used in daily news, financial news, etc. This kind of scrolling text normally moves from left to right in the reading direction, to enable the reader to gain as much information as possible in a period as short as possible, using the smallest possible screen area.
  • Arabic text will generally move from right to left.
  • Asian text will generally move from top to bottom.
  • Fig. 3 An outline of the method used to extract an image of the text in the first region 11 and an image of the text in the second region 12 is given in Fig. 3.
  • video input data representative of the sequence of image frames 8,9,10 representing a sequence of moving images is obtained.
  • the video input data may be in compressed format, e.g. in accordance with international standard ISO/IEC 13818-2.
  • the video input data representative of a sequence of image frames 8-10 representing a sequence of moving images is obtained by separating only the normal video component of the hybrid split screen frames.
  • the method outlined herein is applied to video data representing a sequence of moving images for occupying only a section of an area of display of a screen. Techniques for identifying the boundaries in hybrid split screen video are known as such.
  • step 14 For every frame 8,9,10, all text blocks 11,12 are detected and localised (step 14). Rough rectangular regions 11,12 that contain the text are identified. Any method suitable for this task can be used. Optionally, multiple methods may be used, and the results may be combined. An example of a suitable method is described in US 2003/0021342.
  • step 15 it is determined which regions within the frames 8,9,10 correspond.
  • the first region 11 and the second region 12 are present in each of the frames 8,9,10, as illustrated by the arrows in Fig. 2.
  • each of the regions 11,12 containing text is treated separately, in parallel.
  • step 16 For each of the regions 11,12, the displacement from one frame to the next is estimated (step 16).
  • This step 16 may use one or more of an image correlation technique (Figs. 4 and 5 illustrate examples), video motion vectors (illustrated in Fig. 6) and feature point tracking (illustrated in Fig. 7).
  • a first variant of an image correlation technique these techniques are also known as template matching - involves computing an error measure e between a first image 17 and a second image 18 after geometrically displacing one of them in relation to another by a distance d.
  • the displacement with the lowest error gives the estimated displacement.
  • an exhaustive search within a search space defined by the maximum displacement and a number of different search directions is carried out.
  • an optimisation technique such as gradient descent may be used to make the search more efficient.
  • the error measure e may be the difference between images, possibly limited to the difference between only edge maps computed for each of the images 17,18.
  • one or more horizontal lines 19,20,21 are used to estimate the displacement.
  • a pattern of pixels is chosen and only that pattern is used to estimate the displacement.
  • the patterns may correspond to patterns contained within the regions 11,12 detected and localised in the second step 14. An effect is that any displacement occurring in parts of the frames 8,9,10 not corresponding to scrolling text areas would not influence the displacement estimate. Moreover, the text in the first region 11 is relatively distinctive, which also improves the quality of the displacement estimation.
  • An alternative or additional displacement estimation technique is illustrated in Fig. 6. This technique uses video motion vectors as used, for example in video coding according to international standards ISO/IEC 13818-2 or ISO/IEC 14496.
  • a most prominent motion is selected as the displacement estimation.
  • the modus of the set of motion vectors is selected.
  • a parametric distribution e.g. a Gaussian
  • an algorithm such as RANSAC (RANdom Sample Consensus) is used to find the most profound displacement, viz. the displacement that leads to the minimum number of outliers.
  • RANSAC Random Sample Consensus
  • an unsupervised clustering method may be applied to the motion vectors, and the mean of the largest cluster selected as the displacement estimate.
  • Fig. 7 illustrates another implementation of the step 16 of estimating the displacement of the first region 11 from one image to the next.
  • This embodiment uses landmark, or feature point, tracking.
  • landmark points 23 are detected and matched between two frames 24,25.
  • the displacement of each landmark point 23 can be represented by a motion vector 27.
  • an image 26 with an associated set of motion vectors 27 is obtained, to which the alternative estimation methods outlined above with reference to Fig. 6 can be applied.
  • Landmark tracking is described, amongst others, in Harris, C. and Stephens, M., "A combined corner and edge detector", Fourth Alvey Vision Conference, 1988, pp. 147- 151 and Lowe, D. G., "Object recognition from local scale-invariant features", Proc. Seventh IEEE Int. Conf. on Computer Vision, g, 1999, pp. 1150-1157.
  • the second region 12 is generally static, so that an image of this region can be filtered out (step 28) relatively easily and provided to an OCR application (step 29).
  • the first region 11 At least the copies of the first region 11, but alternatively the whole of the frames 8,9,10, are registered (step 30) such that the first region 11 is generally static from one image to the next. Because the displacement of the first region 11 does not generally correspond to an integer pixel length, this step 30 is preceded by a step 31 of scaling up the image frames 8,9,10 to enhance the resolution of the images represented by them. Registration at pixel accuracy would introduce a small blurring effect in the scroll direction, which may detract from the performance of the subsequent OCR method, in particular where the text comprises characters in a slender font.
  • Either of a single-frame or multi- frame super-resolution technique can be used in this step 31.
  • text is segmented (step 32) from background, based on multiple registered images, in order to provide an image of the text that is thereby filtered from the registered images as output.
  • segmentation is applied anew, but this time on the registered representations of the first region 11 or entire image frames 8,9,10. In the following, the latter option will be used as an example.
  • a simple way of carrying out this step 32 is to blend multiple registered images to obtain a composite image.
  • pixels in at least parts of the composite image corresponding to the area of overlap between the multiple registered images are based on values of the corresponding pixels in the multiple registered images. They could be an average or a sum, for example.
  • the background moves, whereas the text remains static.
  • the composite image has a text region that is sharp and clear and a blurred background.
  • a method such as the one described in US 2003/0021342 is applied directly to the composite image.
  • large background regions may first be segmented, in order to apply the method to only relevant parts of the composite image.
  • a colour segmentation algorithm may be applied to the composite image to identify connected regions similar in at least one of texture and colour.
  • a suitable algorithm is a watershed method, described, for example, in Gonzales, R.C. and Wood, R.E., Digital Image Processing, 2 nd ed., 2002.
  • the criterion incorporates knowledge of the average size of segments that correspond with high probability to regions of text.
  • the size criterion may be used to eliminate certain ones of the regions identified by applying the method of US 2003/0021342 to the connected regions obtained through colour segmentation.
  • the method of US 2003/0021342 involves identifying regions corresponding to at least part of a certain one of the connected regions by calculating a complexity of the region and determining if the complexity of the region exceeds the complexity of at least one other region by a pre-determined ratio.
  • the complexity is calculated by one of a number of different ways.
  • the regions are divided into slices, each encoded into a number of bits and a quantisation scale.
  • the complexity of a region is defined as the sum of the products of the number of bits and the quantisation scale over all slices constituting the region.
  • the regions are encoded in spatial frequency components, and the complexity of a region is defined as the centre of gravity of the coefficients encoding a region.
  • Colour segmentation is good for removing background regions with low texture.
  • a variance in value is computed across a number of the registered images. Pixels having a variance above a certain threshold value are excluded. They are most likely to belong to the (moving) background. Those pixels having a variance below a certain, possibly a lower, threshold, may represent text. Characters typically have a bright body surrounded by dark edges that may or may not contour the entirety of the characters.
  • An algorithm based on this knowledge includes identifying those of the pixels having a variance below a certain threshold and a brightness below a certain level in the number of registered images, and applying a morphological operation for closing contours defined at least partially by the identified pixels.
  • a suitable morphological operation is the binary closing operation as described, for example, in Gonzales, R.C. and Wood, R.E., Digital Image Processing, 2 nd ed., 2002.
  • pixels having a variance below a certain threshold and a brightness above a certain level in the number of registered images are identified. If they lie within any of the closed contours, they are included in the filtered image of the text provided as output.
  • the image of the text is provided as a binary output image, as opposed to a greyscale image, to the OCR application in the final step 29. Due to the quality of the segmentation method, the OCR application can be relatively simple, consuming fewer computational resources.
  • 'Means' as will be apparent to a person skilled in the art, are meant to include any hardware (such as separate or integrated circuits or electronic elements) or software (such as programs or parts of programs) which perform in operation or are designed to perform a specified function, be it solely or in conjunction with other functions, be it in isolation or in co-operation with other elements.
  • 'Computer program' is to be understood to mean any software product stored on a computer-readable medium, such as an optical disk, downloadable via a network, such as the Internet, or marketable in any other manner.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

L'invention porte sur un procédé d'extraction de texte défilant à partir d'une séquence d'images animées, lequel procédé comprend l'obtention de données d'entrée vidéo représentatives d'une séquence d'images fixes (8-10) représentant la séquence d'images animées. Au moins une région (11, 12) contenant du texte à l'intérieur des images de la séquence est détectée. Pour chacune d'une ou plusieurs régions détectées, un déplacement de la région (11) d'une image à une autre est estimé, les images sont enregistrées selon le déplacement estimé de telle sorte que la région (11) est généralement statique d'une image enregistrée à la suivante, et au moins des parties d'images fixes représentant de multiples images enregistrées sont amenées à être utilisées en tant qu'entrée dans un procédé de segmentation de texte à partir d'un fond, basé sur de multiples images de façon à générer une image du texte filtré à partir de celles-ci.
PCT/IB2008/053219 2007-08-16 2008-08-12 Procédé d'extraction de texte défilant à partir d'une séquence d'images animées WO2009022290A2 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
EP07114441 2007-08-16
EP07114441.4 2007-08-16

Publications (2)

Publication Number Publication Date
WO2009022290A2 true WO2009022290A2 (fr) 2009-02-19
WO2009022290A3 WO2009022290A3 (fr) 2009-04-09

Family

ID=40291269

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IB2008/053219 WO2009022290A2 (fr) 2007-08-16 2008-08-12 Procédé d'extraction de texte défilant à partir d'une séquence d'images animées

Country Status (1)

Country Link
WO (1) WO2009022290A2 (fr)

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6937766B1 (en) * 1999-04-15 2005-08-30 MATE—Media Access Technologies Ltd. Method of indexing and searching images of text in video

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6937766B1 (en) * 1999-04-15 2005-08-30 MATE—Media Access Technologies Ltd. Method of indexing and searching images of text in video

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
H. LI: "AUTOMATIC PROCESSING AND ANALYSIS OF TEXT IN DIGITAL VIDEO," UNIV. OF MARYLAND, no. LAMP-TR-59, December 2000 (2000-12), pages 1-123, XP002513699 Maryland, USA *
I. PRATIKAKIS: "Deliverable 2.2. "Semantics extraction from visual content tools - The state of the art" BOOTRAPPING ONTOLOGY EVOLUTION WITH MULTIMEDIA INFORMATION EXTRACTION (BOEMIAE, 20 March 2007 (2007-03-20), pages 1-36, XP002513700 *
JUNG K ET AL: "Text information extraction in images and video: a survey" 1 May 2004 (2004-05-01), PATTERN RECOGNITION, ELSEVIER, GB, PAGE(S) 977 - 997 , XP004496837 ISSN: 0031-3203 the whole document *
LIENHART R ED - ROSENFELD A ET AL: "VIDEO OCR: A SURVEY AND PRACTITIONER'S GUIDE" VIDEO MINING; [KLUWER INTERNATIONAL SERIES IN VIDEO VIDEO COUMPUTING], NORWELL, MA : KLUWER ACADEMIC PUBL, US, 1 January 2003 (2003-01-01), pages 155-184, XP009046500 ISBN: 978-1-4020-7549-0 *

Also Published As

Publication number Publication date
WO2009022290A3 (fr) 2009-04-09

Similar Documents

Publication Publication Date Title
Yin et al. Text detection, tracking and recognition in video: a comprehensive survey
Mosleh et al. Automatic inpainting scheme for video text detection and removal
Crandall et al. Extraction of special effects caption text events from digital video
US20090274364A1 (en) Apparatus and methods for detecting adult videos
EP1840798A1 (fr) Procédé de classement de données d'image numériques
US7840081B2 (en) Methods of representing and analysing images
Barbu Novel automatic video cut detection technique using Gabor filtering
Crandall et al. Robust detection of stylized text events in digital video
Anthimopoulos et al. Multiresolution text detection in video frames
EP2325801A2 (fr) Procédé pour la représentation et l'analyse d'images
Arai et al. Text extraction from TV commercial using blob extraction method
Aghajari et al. A text localization algorithm in color image via new projection profile
Wang et al. Detecting video texts using spatial-temporal wavelet transform
Li et al. An integration text extraction approach in video frame
Jamil et al. Local statistical features for multilingual artificial text detection from video images
Paliwal et al. A survey on various text detection and extraction techniques from videos and images
Raza et al. Multilingual artificial text detection and extraction from still images
WO2009022290A2 (fr) Procédé d'extraction de texte défilant à partir d'une séquence d'images animées
Su et al. A novel model-based segmentation approach to extract caption contents on sports videos
Phan et al. A skeleton-based method for multi-oriented video text detection
Xu et al. Caption Text Extraction Using DCT Feature in MPEG Compressed Video
Anthimopoulos et al. Detecting text in video frames
Elshahaby et al. A system for detection of moving caption text in videos: a news use case
Murthy et al. Robust model for text extraction from complex video inputs based on susan contour detection and fuzzy c means clustering
Covavisaruch et al. Time Stamp Detection and Recognition in Video Frames.

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 08789598

Country of ref document: EP

Kind code of ref document: A2

NENP Non-entry into the national phase

Ref country code: DE

WPC Withdrawal of priority claims after completion of the technical preparations for international publication

Ref document number: 07114441.4

Country of ref document: EP

Date of ref document: 20100204

Free format text: WITHDRAWN AFTER TECHNICAL PREPARATION FINISHED

122 Ep: pct application non-entry in european phase

Ref document number: 08789598

Country of ref document: EP

Kind code of ref document: A2