Bangla OCR :: Literature Study

Paper Review :: aipr98.pdf

2005-08-29T00:42:00.000-07:00

Category: Complete OCR
File Name: aipr98.pdf
Title: A Robust, Language-Independent OCR System

Quick Notes:
Language-independent OCR
Uses HMM to model character
Uses unsupervised adaptation techniques (enhances performance for degraded data, e.g. fax)
Demonstrated language-independence for Arabic, English, Chinese

HMM used in continuous speech recognition (CSR); benefits –
.. language-independent training and recognition methodology …… automatic training on non-segmented data
.. simultaneous segmentation and recognition

Except for the preprocessing and feature-extraction stages, this
OCR system utilizes the BBN BYBLOS continuous speech recognition system without any modification to perform the
training and recognition

Extract features from thin slices of the line image, which is the key to making the recognition system language-independent.

This approach of using HMM departs from other OCR approaches in three ways.
.. First, focused on the problem of language-independent recognition: the major components of the system (feature extraction, training, and recognition) are designed to be script-independent.
.. Second, training and recognition are performed using an existing continuous speech recognition system with no modification except of course for preprocessing and feature extraction.
.. Third, there is no need to perform any pre-segmentation either at the character or at the word level.

A segmentation-free approach is important for the recognition of degraded documents where characters are often connected, and makes it easy to train the OCR system on new corpora and to apply the system to new scripts.

This OCR system includes two parts:
1. the training system
2. the recognition system

Paper Review :: character_recog_eng.pdf

2005-08-29T00:02:00.000-07:00

Category: Overview
File Name: character_recog_eng.pdf
Title: Character recognition Systems for the non-expert

Comments: good for gathering basic ocr knowledge

Quick Notes:
Character recognition systems (CRS) are subsets of pattern recognition systems.

[Patterns / Words]

Input may come from – online or offline devices.

--> CRS components

a] Pre-processing functions
.. noise reduction
.. skeletonisation (thinning)
.. normalization
...... moment invariant techniques
...... fourier descriptions
...... boundary-based techniques
...... vector analysis
.. segmentation

b] accepts pre-processed inputs &
.. extracts characteristic features (feature extraction maps the whole of each input pattern from its original Euclidean spatial system of co-ordinates onto a single point in a feature space. Feature space is defined by N extracted features, i.e. has N dimensions)
{loop, end-points, dot, junction, contour}
...... horizontal & vertical histograms
...... curvature info (slopes) & local extrema of curvature (of line making-up/fitting a word)
...... topological features (loops, dots, junctions)
...... parameters of polynomial (or other) curve fitting functions
...... contour info

c] classification component (e.g. neural net)
.. assign a label to pattern
techniques:
...... rule-based systems
...... decision trees
...... clustering techniques
...... artificial neural networks
...... hidden Markov models

d] post-processing
.. verification (increase level of confidence in classification made)
.. action execution
.. adaptation (reduce gaps between expected & actual performance)
...... ANN (alter own weights)
...... HMM (probabilistic parameters)

Segmentation

2005-07-13T22:43:00.000-07:00

[~] Skew angle estimation

source: project OCR by group algfk-21 at Chalmers

A document obtained by a digitizer may exhibit various kinds of geometrical distorsions. For instance if the document is not properly aligned on the scanner the document could become skewed (rotated). This problem can however be corrected by estimating the skew angle.

One way of doing this is by hypothesis-testing --
- projec the image along a number of axes
- compute the orientation-dependante histograms
- look for the direction that maximizes an alignement criterion A(a)
- Finally, the estimated skew angle is given as the angle that maximizes A(a)

[~] Character extraction

source: project OCR by group algfk-21 at Chalmers

X-Y-Tree Decomposition / Iterative Projection Profile Cuttings method:

basic idea is that the document contains vertical and/or horizontal structure. Often documents contain blocks of text that contains rows of characters. The basic operation of this algorithm is a projection of the document image on a horizontal or vertical axis.

-1 Compute the horizontal projection of the entire page
-2 Analyse the projection profile to extract the lines
-3 For each line, compute the vertical projection profile
-4 Analyse the projection profile obtained in step 3 to extract the characters

In some cases two projections is not enough, as for more complex documents

This segmentation algorithm is simple and have problems ---
with characters that is disjoint, these characters can get segmented into two different characters
On the other hand if two characters are connected they will be segmented into one block

Image pre-processing

2005-07-13T22:14:00.001-07:00

[~] Salt and Pepper noise removal

black dots in the image --> Pepper
holes in the image --> Salt

source: project OCR by group algfk-21 at Chalmers

The method is based on two basic binary image morphological operations:
1] Dilations --> (expansion, the act of expanding an aperture)
2] Erosions --> (wearing away, the mechanical process of wearing or grinding something down)
(based on the Minkowski addition and subtraction)

open and close operation:
- Opening an image will eliminate small islands, sharp peaks and thin lines.
- Closing an image will fuse narrow breaks, close small holes and smooth contours.

[~] Character Boundary Generation

source: project OCR by group algfk-21 at Chalmers

- finding the boundary of a character-image
- method starts with a gray-scale image and ends up with the characters boundary that is one pixel wide

... two steps:
step-1: edge detection
- performed by scanning the whole grey-scale image and picking out the boundary-pixels.
- A boundary-pixel is defined as a pixel whose grey-level is above the average gray-level and has a neighbor whose level is below the average grey-level.
- boundary generated by this algorithm has a width of maximum two pixels.
step-2: edge thinning
done in two passes:
:: In the first pass
- the boundary-image is scanned
- all pixels with three or five neighbors which are not branching are deleted.
- repeated until no deletions occur.
:: In the second pass
- all diagonal lines are thinned.

To get an even better result, the grey-scale image should undergo some form of enhancement to filter out noise and to sharpen the image before the boundary is generated.

OCR: 3 primary steps

2005-07-13T21:30:00.001-07:00

Hey guyz, this is a summary of what i thought was useful from the following site ..

source: project OCR by group algfk-21 at Chalmers

1] representation process

doc scanned --> got an image --> process image to achieve a higher level form of data

:: image pre-processing
- filtering out noise and increasing the contrast
:: segmentation
- separate the characters from each other
i.e. whole bunch of smaller images each representing a character
:: feature extraction
- raw digitized data is then mapped to a higher level by extracting special characteristics and patterns of the image
- can make the image invariant to rotation, translation, scaling, line-thickness etc
- could also remove redundant information to compress the data amount
- higher level image is then stored in some special way --> perhaps in a vector, tree or a graph

2] learning step

generation of a known database containing the high level representation of all known characters
(done in the same way as in the representation process)

- system learns to separate the classes
- Usual forms of database --> matrices, trees, graphs, weights in the net achieved by training (neural networks)

3] identification / classification process

classifies the unknown character given its high level representation & the information learned from the known database

:: identification approaches
- statistical approach
based on a similarity measure that is expressed in terms of a distance measure or a discriminant function
- syntactical approach
based on parsing syntax trees in the database or recognition of pattern grammars
- Neural Network
more of a fuzzy classifier using trained weights