Kannada Text Image Matching (KTIM) Algorithm

Kannada Text Image Matching (KTIM) Algorithm

Mrs.Tanuja K, ISE Dept, Mrs. UshaKumari, ISE Dept,

Acharya Institute of Technology, Bangalore 90 Acharya Institute of Technology, Bangalore – 90

Email: ushavekappa@gmail.com

Abstract In this paper we present an algorithm to retrieve Kannada text image from huge collection of Kannada image using the Kannada text matching algorithm based on BLPOC (Band – limited phase only correlation) function which provides the higher correlation peak and better discrimination capability. This algorithm can achieve the effective and efficient search and retrieval from a huge collection of Kannada images by matching image at word level and not at character level . On the given query image Kannada text matching algorithm based on BPLOC function is used to match the words and retrieves the relevant document.

Index Terms

Morphological operation, Phase only correlation, Band

• limited phase only correlation based image matching

  1. INTRODUCTION

     A number of approaches have been proposed in recent years for efficient search and retrieval of document images. The collections of historical prints and books exist in Indian languages that need search options in images. The document images of such collections cannot be recognized accurately. There is a need for easy and efficient access to such collections. The search procedures available for text domain can be applied, if these document images are converted into textual representations using recognizers. However, it is an infeasible solution due to the unavailability of efficient and robust OCRs for Indian languages.

     Addressing this problem, this paper proposes an efficient document image retrieval algorithm using Band limited phase only correlation based image matching (BPLOC) an image matching technique using the phase components in Inverse Fast Fourier Transformation which determines the phase angle of input image and query image that helps in matching word for the retrieval of document. This approach is faster as it does not match the image pixel by pixel.

     The remainder of the paper describes our current development effort in more detail. Section III describes the architecture of the research we are developing. Section IV details the implementation procedures we are developing. Section V presents some experimental results. Section VI concludes the paper.

  2. SYSTEM ARCHITECTURE

     Fig(1) System Architecture

     This system accepts a text query from users. The text query is initially converted to an image by word rendering, features are extracted from these images and then a search is carried out for retrieval of exact documents. Results of the search are pages from document image collections containing the retrieved words sorted based on their relevance to the query. This work mainly aims at addressing some of the issues involved in effective and efficient retrieval in document images with effective representations of the word images.

  3. IMPLEMENTATION

     An efficient and effective mechanism for retrieval of

  4. MATCHING TECHNIQUE

In this section, the principle of phase-based image matching using the limited Phase-Only Correlation (POC) function (which is sometimes called the phase-correlation function)is explained . Consider two N1×N2 images, f(n1, n2) and g(n1, n2), where we assume that the index ranges are n1 = M1 · · ·M1 (M1

> 0) and n2 = M2 · ··M2 (M2 > 0) for mathematical simplicity, and hence N1 = 2M1 + 1 and N2 = 2M2 +

1. Let F(k1, k2) and G(k1, k2) denote the Fast Fourier Transform(FFT) of the two images. F(k1, k2) is given

by

relevant Kannada document from a large Kannada

F(k , k ) f (n , n

)W k1n1W k2n2

document image collection is presented in this work. This involves three phases: first phase includes pre processing, which is preparing the source image for searching the query Kannada word by removing the

1 2

n1,n2

= AF (k1 , k2

1 2 N1 N2

)e j F ( k1 ,k2 )

noise, second phase includes generating the texted where k1 = M1 · · ·M1, k2 = M2 · · ·M2,

query image from the query word, and third phase includes matching of images to find the query word in the source image

W e

N

N

1

j 2

N

N

N1 W

1 2

1 2

, 2

j 2

e N2

and

n ,n

In first phase, consider a source image, which is generally in form of RGB is converted to the grey

denotes

M1

n1M1

M 2

n2 M 2

AF (k1, k2) is

scale image. Then, this grey scale image is, in turn, converted to the binary image, i This process of conversion will help to performing morphological

amplitude and F (k1, k2) is phase. G(k1, k2) is defined in the same way. The cross-phase spectrum RFG (k1, k2) is given by

operation. Morphological operation is performed to

F (k , k )G(k , k )

initiate the dilation, which helps in differentiating two

R (k , k

) 1 2 1 2

words delimited by a space. Then, fill the holes to find any picture in the image and remove the big area, which might be, say, a photograph. Then, record the

FG 1

2

F (k1

, k2

)G(k1

, k2 )

coordinates, height and width of each word in image document.

Second phase includes generating the query image from the query word. Convert the query image to binary image. This helps in comparison of input image with the query image.

j ( k1 ,k2 )

e

e

=

where G(k1, k2) is the complex conjugate of G(k1, k2) and (k1, k2) denotes the phase difference F (k1, k2)G(k1, k2). The POC function rfg(n1, n2) is the Inverse Fast Fourier Transform ( IFFT) of RFG(k1, k2) and is given by

Third phase includes matching of images to

r (n , n

) 1 R

(k , k

)W k n W k n

find the query word in the source image. The method used is KTIM algorithm. This phase has input as two

fg 1 2

N1 N

FG 1 2

2 k1 ,k2

1 1 2 2

N1 N2

images, source image and query image, which are converted into binary form.

where

k ,k

denotes

M1

k1M1

M 2

k2 M 2

1 2

1 2

When two images are similar, their POC function gives a

distinct sharp peak. When two images are not similar, the peak drops significantly. The height of the peak gives a good similarity measure for image matching, and the location of the peak shows the translational displacement between the images.

Band-Limited Phase-Only Correlation (BLPOC)

In the POC-based image matching method, all the frequency components are involved. However, high frequency tends to emphasize detail information and can be prone to noise. To

1. Calculate the POC function

   1 2 1 2 1 2f g

   1 2 1 2 1 2f g

   M1M2 (n , n ) between f (n ,n ) and g(n ,n );

2. Calculate =arg to select the rotation-normalized image (n1,n2);

3. Estimate the image displacements( , ) between (n1,n2) and g(n1,n2) from peak

   M1M2

   location of (n , n ) ;

   eliminate meaningless high frequency components, K. Ito et

   f g 1 2

   al. [9] proposed the Band-Limited Phase-Only Correlation (BLPOC). The BLPOC limits the range of the spectrum of the given image. Assume that the ranges of the inherent frequency band of are given by u=-U0,U0 and v=- V0,V0, where 0<=U0<=M0, 0<=V0<=N0. Thus, the effective size of spectrum is given by L1=2U0+1 and L2=2V0+1. The BLPOC function is defined as

4. Extend the size of (n1,n2) and g(n1,n2) by and pixels for n1 and n2 direction to both f`( n1,n2) and g`( n1,n2);

5. Extract the effective the text regions f(n1,n2),

   g(n1,n2) from f`( n1,n2) and g`( n1,n2);

6. Detect the inherent frequency band (K1,K2) from the 2D DFT of f(n1,n2);

   p

   p

   U0 v

   1 U0 V0

   j 2( mu nv

   0 m, n

   gf

   L L

   L L

   uU

   vV

   RGF

   (u, v)e

   L1 L2

7. Calculate the band limited POC

   K K

   1 2 0 0

   1 2

   function(BLPOC) (n , n ) ;

   f "g" 1 2

   where m=-U0,,U0 and n=-V0,,V0. When two images are similar, their BLPOC function gives a distinct sharp peak. Also, the translational displacement between the two images can be estimated by the correlation peak position. Experiments indicate that the BLPOC function provides a much higher discrimination capability than the original POC function

   Algorithm for Kannada text Image matching: Procedure : Word matching using BPLOC function Input:

   f(n1,n2): the registered kannada text image g(n1,n2): the kannada text image need to be verified Output:

   Matching score between f(n1,n2) and g(n1,n2);

   1. Begin

   2. Store in advance a set of rotated images f(n1,n2) of f(n1,n2) over the angular range

maxmax with an angle spacing 10;

1. Compute the matching

   P

   P

   score; SK1K2 [ f ", g"]

2. end

   1. EXPERIMENTAL RESULTS

      Figure (5.1) shows the input to the system

      Figure (5.2) Query to the system

      Figure (5.3) Shows the output

      Since we are using database approach for the character recognition, in this approach for each character we need to have details like Character name, Character BMP image. This takes lot of space as well as lot of computation involved in recognizing the character.

   2. CONCLUSION

      In this paper we have presented retrieval in large document image collections. The matching technique is based on phase – based image matching, for search in large collections of document word images is applied to obtain good performance. The approaches used for word spotting so far, dynamic time warping and/or nearest neighbor search tend to be slow for large collection of books. Direct matching of pixels in images is inefficient due to the complexity of matching and thus impractical for large

      databases. This problem is solved by directly storing word image representations.

      Some of the possible directions in which the future work can be carried out are as below. The effect of combination of different fonts in a single collection can be one possible direction for exploring the feasibility of the proposed technique and improving it. Multi-lingual document images are not handled in the proposed technique. Segmentation is not handled in this paper

   3. Reference

1. Koichi Ito, Ayumi Morita, Takafumi Aoki, Tatsuo Higuchi, Hiroshi Nakajima, and Koji Kobayashi, A Fingerprint Recognition Algorithm Using Phase-Based Image Matching for Low-Quality Fingerprints

2. Ashwin T V 2000 A font and size independent OCR for printed Kannada using SVM. M E Project Report, Dept. Electrical Engg., Indian Institute of Science, Bangalore

3. A. Balasubramanian, Million Meshesha, and C.V. Jawahar Retrieval from Document Image

   Collections

   Centre for Visual Information Technology, International Institute of Information Technology,

   Hyderabad – 500 032, India

4. Ito, K., Nakajima, H., Kobayashi, K., Aoki, T., Higuchi, T.: A fingerprint matching algorithm using phase-only correlation. IEICE Trans. Fundamentals E87-A(3), 682691 (2004)