[attachment=9687]
1. INTRODUCTION
As computer, compress technology, storage media and high speed communication skill are developed dramatically; digital video has become one of the most important elements in many applications such as education, news and games. Multimedia data are also getting bigger than before. In order to extract and search important information from a huge amount of video data, we need to extract text from video. Text is obviously an important element in video. So extracting text appears as a key clue for understanding contents of video and for instance for classifying automatically some videos. Videotext detection and recognition has been identified as one of the key components for the video retrieval and analysis system. Videotext detection and recognition can be used in many applications, such as semantic video indexing, summarization, video surveillance and security, multilingual video information access, etc.
Videotext can be classified into two broad categories: Graphic text and scene text. Graphic text or text overlay is the videotext added mechanically by video editors, examples include the news/sports video caption, movie credits etc. Scene texts are the videotexts embedded in the real-world objects or scenes, examples include street name, car license number, and the number/name on the back of a soccer player. This report is to address the problem of accurately detecting and extracting the graph videotexts for videotext recognition. Although the overlay text is manually added into the video, the experiments showed they are even as hard to extract as many video objects, such as face, people etc. This is due to the following reasons: 1. Many overlay texts present in the cluttered scene background; 2. There is no consistent color distribution for texts in different videos. Consequently, the color-tone based approach widely used in face or people detection application actually cannot be applied in text detection. 3. The size of the text regions may be very small such that when the color segmentation based approach is applied, the small text region may merge into the large non-text regions in its vicinity.
Here we used edge detection based method for extracting the text and it is implemented using Matlab. Here the two critical angles are defines and the text is extracted and recognized using the coincidence of the edges of the image with the threshold defined based on the critical angles.
The two angles used here are 180 degree and 90 degree. If we add more threshold angles the accuracy of the extracted text can be increased but the image fragments also may get into the final result as the coincidence of the are edges.
2. MAIN CONCEPT
Text extraction in video consists in three steps. The first one is to find text region in original images. Then the text needs to be separated from background. And finally a binary image has to be produced (for example, text is white and background is black)
Difficulties of such a project can be classified in following main categories:
1. Background and text may be ambiguous.
2. Text color may change: text can have arbitrary and non-uniform color.
3. Background and text are sometimes reversed.
4. Text may move.
5. Unknown text size, position, orientation, and layout: captions lack the
structure usually associated with documents.
6. Unconstrained background: the background can have colors similar to the
text color. The background may include streaks that appear very similar to
character strokes.
7. Color bleeding: lossy video compression may cause colors to run together.
8. Low contrast: low bit-rate video compression can cause loss of contrast
between character strokes and the background.
3. CHARACTERISTICS OF TEXT IN VIDEO
In order to have good result in text extraction, it is necessary to deal with the characteristics of text. Most of the time, artificial text in video has the following properties:
- Monochrome: that is to say that text consist of same or similar colors
- Easily readable for human: that is to say that text is distinguishable with
background and maintains a readable size (not too big and not too small)
- Appearing in multiple subsequent frames: that is to say that text is fixed or move
linearly at a low speed and maintains same size and shape.
We made the following assumptions for doing this particular project:
- No real-time processing
- Monochromatic letters
- Non-moving text
- Size (height and width) restricted text