Text%20Detection%20in%20Video

1
Text Detection in Video

• Min Cai
• 2002.3.13

2
Background

• Video OCR Text detection, extraction and
  recognition
• Detection Target Artificial text
• Text detection
• Detect the region from Single frame
• Refine the region by combining consecutive frames

3
Existing Work
Feature Extraction Text Detection based on feature
Color Connected-component
Texture Texture-Segmentation
Edge Top-Down Bottom-Up
4
Connected-component-based methods

• Basic idea
• Treat text as an uniform color (color level) and
  classify each pixel as text or non-text according
  to the color value.
• Combine connected text-pixels into connected
  components.
• Group collinear connected components into a text
  string.
• Advantage
• Can detect an arbitrary orientation text ----
  with similar color and in a simple background.
• Disadvantage
• Sensitive to color variance
• Lossy compression of video introduces color
  bleeding
• Complex background

5
Texture Segmentation method

• Basic idea
• Treat text as a type of texture
• Use texture segmentation algorithms to detect
  text
• Gabor Filter
• Gaussian derivatives
• Advantage
• Can segment text areas graphic areas in a
  simple background efficiently. It is usually used
  in document analysis.
• Disadvantage
• Time-consuming
• Cannot handle well a text embedded in various
  background.

6
Bottom-Up method

• Basic idea
• A seed region is defined as a small region with
  high edge density.
• Grow a seed region into successively larger
  components until all seed regions are reached on
  the image.
• Advantage
• It is a generic method to detect a homogeneous
  object of various shape. That is, it can detect
  not only a rectangular object, but also other
  shapes.
• Disadvantage
• Sensitive to noise.
• Can not handle the large range of font-size.
• Sensitive to the stroke density (different
  language).

7
Top-Down method

• Basic idea
• Based on run-length smoothing algorithm
• Analyze horizontal and vertical projection
  profiles
• Advantage
• Can detect the boundary of horizontal alignment
  text string quickly and correctly
• Noise insensitive
• Disadvantage
• Cannot handle diagonal alignment text.
• One pass of horizontal vertical projection
  cannot handle the complex layout.

8
Analysis (1)

• A certain contrast against background
• Artificial text strings are designed to be read
  easily
• A certain stroke density
• Text strings always appear horizontally
• Spatial cohesion
• Characters of the same text string are of similar
  heights, orientation and spacing
• Size constraint
• Text strings have certain size restriction
• A text string appears in multiple consecutive
  frames and the similar position.

9
Analysis (2)
Problems Resolutions
How to extract more useful edge? Local Thresholding
How to highlight text areas? Text area recovery
How to detect text regions fast and correctly? Coarse-To-Fine detection
10
Single Threshold
11
Local threshold (1)

• Use a small kernel (red) to scan the whole image.
• In a bigger window (gray) surrounding the kernel,
  calculate the local threshold corresponding to
  its local histogram.

b. Local threshold selection
a. Window move
12
Local threshold (2)
13
Text-like area recovery (1)
Before recovery After recovery
14
Text-like area recovery (2)
Before recovery After recovery
15
High pass filter
16
Coarse-to-Fine detection

• Using Top-down scheme to detect text-like areas

17
Detect text-like areas
1)
2)
3)
4)
b. Coarse vertical projection
18
Refinement

• Combine the neighboring text areas with similar
  height
• Using size constraints to remove unsatisfied areas

19
Multi-frame analysis

• Text region matching
• Find all the regions corresponding to the same
  text
• Text region enhancement
• Enhance the text image quality by multi-frame
  integration
• Repetitive text elimination
• Only record the text at its first emergence.

20
Thank you!
End