A HYBRID TRADEMARK RETRIEVAL SYSTEM USING FOUR-GRAY-LEVEL ZERNIKE

MOMENTS &

IMAGE COMPACTNESS INDICES

Department of Mechanical & Electro-Mechanical Engineering

Innchyn HER, Yung-Han HSIAO and Huan-Kai HUNG

CONTENT Introduction Objectives Trademark

Methodology Smallest Enclosing Circle Compactness Indices Four Level Images Exclusive Trademark features Weighting and Normalization

Results Conclusion Reviewer Comments

INTRODUCTION

Still relies largely on manual and tag-based methods

Using search codes risks the danger of inaccuracy,

inconsistency, and inefficiency.

Content-Based Image Retrieval (CBIR) Contour-based approach

Region-based techniques

Zernike moments for trademark retrieval schemes

OBJECTIVES

Propose a hybrid retrieval system for trademarks.

Use Zernike moments for their resistance to noise, their

invariance to rotation, and multi-resolution capabilities.

Two new image features are proposed; Four-Gray-Level and

Compactness Indices.

In this system humans set up a benchmark and result of

Hybrid system will compare with benchmark.

TRADEMARKS A trademark is a mark that identifies one person's goods There are four different kinds of trademarks

(a)Word trademarks (b)Graph trademarks (c)Sign trademarks (d)Combined trademarks

METHODOLOGYSmallest Enclosing Circle

Zernike moments has property of rotational invariance. To achieve

scaling and translational invariance, additional arrangements are

needed.

Since Zernike polynomials are defined within the unit circle, a method

for conveniently finding the smallest enclosing circle for a trademark is

necessary.

Propose a simplified version, based upon the Berg et al. recursive

formula, for locating the smallest enclosing circle for a trademark

image.

SMALLEST ENCLOSING CIRCLE

Firstly, extract digitally its contour, then, search among the boundary points and find the two points that are most distantly apart, say, points a and b. Construct circle D1

Locate point c, which is farthest from the center O1. Construct circle D2 by points a, b, and c.

SMALLEST ENCLOSING CIRCLE

Lastly, if there still are some

remaining points outside D2, there is

a final modification to make. Find

point d that is most distant from O2.

Let be the length of the

new diameter, construct D3, which is

the smallest enclosing circle used in

this paper.

2 2O d O c

WRAP CONTOUR Besides Zernike moments, a region-based method we need some

contour-based features that can represent the gross shape. Propose a wrap contour concept.

(b) (c)

(a)

Using contours as features. (a) A trademark and its smallest enclosing circle, (b) The convex hull for that trademark, (c) The wrap contour for that trademark.

Locate smallest enclosing circle.Radial lines are drawn The intersections are recorded as a function r()If there is no intersection, r=0Finally, let the wrap contour r() become a closed and continuous curve in the polar plane

COMPACTNESS INDICES

The wrap contour used to define two new features called

image compactness indices, CI1 and CI2. They can be regarded

as mixed contour-based and region-based features, and are

defined as:

Compactness index 1 (CI1) =

Compactness index 2 (CI2) =

Areaof Wrap ContourAreaof Smallest Enclosing Circle

ImageAreaAreaof Wrap Contour

FOUR-LEVEL IMAGES For simple images as trademarks, some researchers suggested

using only a two-level (binary) scale. Transforming trademarks into binary images sometimes incurs significant losses in the features.

Losing boundaries from using too few gray levels.

(a)Some internal boundaries are missing,

(b) The plus sign is gone.

FOUR-LEVEL IMAGES

We propose using four gray levels that correspond to levels 0, 85, 170,

and 255, in a 0 to 255 grayscale.

A trademark image is then transformed, with evenly spaced threshold

values, to an image that contains only these four gray levels.

By thus using more than two gray values, the loss of image features is

effectively reduced.

EXPERIMENT AND RESULTS In this paper, 2020 trademark images were collected and

compiled to form the experimental database. These Table used as benchmarks to tested hybrid scheme.

Human-Picked Groups of

Similar Trademark Patterns

Nike

Goblet

Recycle

Michelin

Monogram

EXPERIMENT AND RESULTS A collection of eight weighting factors w1: compare the whole input image with whole images from the

database. w2: compare core images only. w3: compare the core of the input image with the whole images

from the database. This is to remove the frame of the input image. w4: compare the inversed core of the input with the whole database w5: compare CI1's of whole images.

w6: compare CI2's of whole images. Lastly, w7 and w8: compare CI1's and CI2's of the core images, respectively.

j 1 2 8w {w ,w , w }

EXPERIMENT AND RESULTS Using different gray-levels with Zernike moments. (a) Two-level scheme

(b) Four-level scheme

Input image

Input image

Gathers more monogram-like images in its top rows

EXPERIMENT AND RESULTS

(a) Hu's moment invariants

(b) Traditional Zernike moments

(c) Kim's modified Zernike

(d) Our compactness index

CI1

w5=1 (compare CI1's of whole images)

EXPERIMENT AND RESULTSBy using more image features to improve the search efficiency

(a) Hu's moment invariants

(b) The original Zernike,

(c) Kim's modified Zernike

EXPERIMENT AND RESULTS

(d) Our four-level Zernike

(e) Our Compactness indices

(f) Combining double Zernike and Compactness indices

{1,4,0,0,0,0,0,0}

{0,0,0,0,1,0,4,1}

{1,4,0,0,1,0,4,1},

PERFORMANCE SUMMARY

In average, both our double Zernike schemes and CI-only schemes got roughly 60% of human-picked images in image listings.

This was no better than Kim's modified Zernike (65% in average) and more or less equivalent to traditional Zernike (58% in average).

But Combination of double-Zernike and CI's yielded very good (95% in average) results.

A 64×64 image size was used for the input and an input image must go through a series of preprocessing processes.

All image features are then extracted, paired with their weighting factors and compared with stored data in the database.

CONCLUSIONBoth region-based image features (i.e., four-gray-level Zernike

moments) and contour-base features (the proposed compactness indices) were used.

A simplified method for finding the smallest enclosing circle was also presented.

Features of different categories were conjoined via a weighting scheme.

Experiments have verified that, when human viewpoints were used as standards, this hybrid scheme performed considerably better than some existing retrieval methods.

REVIEWER COMMENTS Weighting factors are not clear. Experiments needs better organization and more comparison.

Result comparison with table. Proposed method should compare with recent work. The

literature review is too limited to trademark retrieval (its an image processing paper).

What kind of shape can not be handled well? More justification needed? With some images.

Smallest enclosing circle algorithm is not complete. No proof of the validity. Empirical method.

REFERENCES H. Freeman, "On the Encoding of Arbitrary Geometric Configurations," IRE Trans. Electronic

Computing, vol. 10, pp. 260-268, 1961. H. L. Peng and S. Y. Chen, "Trademark Shape Recognition Using Closed Contours", Pattern

Recognition Letters, vol. 18, pp. 791-803, 1997. C. C. Huang and I. Her, "Homomorphic Graph Matching of Articulated Objects by An Integrated

Recognition Scheme," Expert Systems with Applications, vol. 31, pp. 116-129, 2006. P. Y. Yin and C. C. Yeh, "Content-based Retrieval from Trademark Databases," Pattern Recognition

Letters, vol. 23, pp. 113-126, 2002. M. K. Hu, “Visual Pattern Recognition by Moment Invariants,” IRE Trans. Information Theory, pp.

179-187, 1962. G. Ciocca and R. Schettini, “Content-based Similarity Retrieval of Trademarks Using Relevance

Feedback,” Pattern Recognition, vol. 34, pp. 1639-1655, 2001. N. K. Kamila, S. Mahapatra and S. Nanda, "Invariance Image Analysis Using Modified Zernike

Moments," Pattern Recognition Letters, vol. 26, pp. 747-753, 2005.

THANK YOU!