In text detection, our
previously proposed algorithms are applied to obtain text regions
from scene image. First, we design a discriminative character
descriptor by combining several state-of-the-art feature detectors
and descriptors. Second, we model character structure at each
character class by designing stroke configuration maps.
Roadmap to Membership of RICS - Pathways and Routes
SCENE TEXT RECOGNITION IN MOBILE APPLICATION BY CHARACTER DESCRIPTOR AND STRUCTURE CONFIGURATION
1. SCENE TEXT RECOGNITION IN MOBILE
APPLICATION BY CHARACTER
DESCRIPTOR AND STRUCTURE
COMFIGURATION
CHERIYAN K M
2. INTRODUCING…..
Valuable information form an image.
To extract an information.
Automatic and Effective scene text detection.
Recognition algorithm.
Factors affecting on extraction.
Cluttered background.
Difference in text pattern.
Difficult to model the structure of character.
Lake of discriminative pixel level appearance.
Structure features from non-text background outliers.
Different word , may diff. characters , in various fonts ,
styles and size.
3. Two activities;
Text detection.
Localize the image region containing the text characters.
Based on
Color uniformity and
Horizontal alignment of text char.
Text recognition.
Transform pixel-based text into reliable codes.
Distinguish diff. text characters , Properly compose the text
word.
62 identity category of text characters.
9 (0-9)
26 (a-z)
26 (A-Z)
Two schemes;
Character recognizer to predict the category of text char.
Binary character classifier to predict the existence of ctgry.
4. RELATED WORKS
Optical Character Recognizer (OCR) system.
Many algorithms are proposed;
Weinmen:- combined the Gabor-based appearance
model.
Neumann:- based on extremal region.
Smith:- based on SIFT.
Mishra:- adopted conditional random field.
Lu:- modeled the inner character structure.
Coates:- extracted local features of character patches.
5.
6. LAYOUT BASED SCENE TEXT DETECTION
A text;
Instruction
Identifier
Uniform color
Aligned arrangement
Two processes are employed to complete layout
analysis
1. Color Decomposition
2. Horizontal Alignment
Improved to compatible with mobile app
7. LAYOUT ANALYSIS OF COLOR
DECOMPOSITION
Boundary clustering algorithm base on bigram color
uniformity.
Group pixels of same color into a layer.
Character boundary boarder b/w txt and bg.(color
pair)
Create a vector of color pair (txt and bg).
8.
9. LAYOUT ANALYSIS OF HORIZONTAL
ALIGNMENT
Text information(string)
Several
character
members
In similar
size
Approximately
horizontal
alignment
The geometrical properties to detect the existence of text characters
12. STRUCTURE BASED SCENE TEXT
RECOGNITION
To extract text information.
Binary classification problem.
Character classes(Queried characters).
Binary classifier:- to distinguish character class
from other classes or bg outliers.
Eg: Character class A predict patches containing A as
positive. And other as negative.
Two activities;
1. Character descriptor.
2. Stroke configuration.
13. CHARACTER DESCRIPTOR
Extract structure features.
4 different key points features;
1. Harris Detector:- To extract Key points from corner and
junction.
2. MSER Detector:-To extract Key point from stroke
component.
3. Dense Detector:- To extract Key point uniformly.
4. Random Detector:- To extract the preset number of
Key points in a random pattern.
14.
15. Flowchart of our proposed character descriptor
HOG:-features are Calculated as observed feature vector x.
(Histogram of Oriented Gradient)
•Selected as local feature descriptor( compatibility
with all 4 key point detectors).
16. SIFT and SURF are not employed
Normalization of character patches(128x128).
Feature Quantization: to aggregate the extracted
features
Bag-of-Words(BOW) Medel:- Applied to key points from
all 4 feature detector.
Gaussian Mixture Model(GMM):Applied to key points
from DD & RD.(fixed number and location of key point)
Now mapping both into characteristic Histogram as
feature representation.
Cascading BOW and GMM-based feature repr. ,we
get Character Descriptor.
17. CHARACTER STROKE CONFIGURATION
Stroke:- Region bounded by two parallel boundary
segments.
Stroke width
Stroke orientation
Characters are connected strokes with
configuration.
Structure Map of Strokes is stroke configuration.(
is consistant)
Eg: B have 1 vertical stroke
2 arc strokes.
B
18. Synthesized character generator: Estimate stroke
configuration from computer s/w.(Provide accurate
skeleton and boundary)
Synthetic font training dataset(20000 are selected
out off 67400 character patches)
Contain 62 class of characters(128x128 pixel)
Compose Stroke Configuration
Step1
Discrete Contour Evaluation(DCE):obtain boundary and
skeleton. Skeleton pruning on the basis of DCE.
DCE simplifies the character(using polygon and small
no. of vertices)
DCE and Skeleton pruning are invariant to deformation
and scaling.
19. Step2
Estimate stroke width and orientation
Width: length along normal
Orientation: tangent
Sampling from character boundary
128 samples.
So that no. of samples = length
Estimating
Taking two neighboring sample point to fit a line.
Approximately collinear.
A quadratic curve.
20. Step3
Calculate Skeleton-based stroke map
Consistency of stroke width and orientation.
Width no larger than 3 Orientation no larger than /8
Construct stroke section: If sample point satisfying
the stroke related features.
Construct junction sections: If they are not.
Skeleton points are extracted.
21.
22. STROKE ALIGNMENT METHOD
To handle various fonts, styles …..etc
Mean value of all stroke configuration.
Mean value,
23. D=Distance b/w stroke configurations of two
samples
S=Mean value of stroke configurations.
Ti=Transformations applied on strokes of i-th stroke
configuration.
g(Ti)=Amplitude of the transformation.