• No results found

A Novel Shape Feature Extraction Technique for Content Based Image Retrieval (CBIR) Systems

N/A
N/A
Protected

Academic year: 2022

Share "A Novel Shape Feature Extraction Technique for Content Based Image Retrieval (CBIR) Systems"

Copied!
12
0
0

Loading.... (view fulltext now)

Full text

(1)

A Novel Shape Feature Extraction Technique for Content Based Image Retrieval (CBIR) Systems

K.Geetha1, D.Yuvaraj2, M.Sivaram3, A. Mohamed Uvaze Ahamed4

1Department of Computer Science and Engineering, M.I.E.T Engineering College, Trichy-7, Tamilnadu, India

2Department of computer science,Cihan University - Duhok, Kurdistan Region- Iraq

3Department of Computer Networking, Lebanese French University -Erbil, Kurdistan Region- Iraq

4Department of Computer Science, Cihan University - Erbil, Kurdistan Region, Iraq

Abstract

Content-Based Image Retrieval is regularly developing examination topic into the subject in regards to pc innovative and far sighted because of the reality its appropriateness inside different utility areas. This article gives form essentially based structure descriptor for CBIR frameworks. The essential advances concerned are image preprocessing, point division, shape trademark extraction or closeness estimation. The development 3 stages are normal in view of on line (question image) and disconnected (dataset image) handling. The picture preprocessing base performs photograph resizing, zero cushioning, gaps expulsion, at that point labyrinth evacuation. The perspective comparable division calculation is back in congruity with concentrate target confine pixels. The component extraction footsie performs rise to scale standardization (EDN) and even place standardization (EAN). The components present by utilizing EDN are utilized so notice file point since sum region list focuses calculation. In EAN the form focuses are gotten through getting the triangle inside two abutting purposes of shape or centroid. The proposed work performs EAN by utilizing applying region about the fragment procedure as a substitute of region on the triangle technique, at that point that does not lose shape measurements at someone cost. It is more noteworthy right as per standardize the form focuses in light of the fact that understanding measure. Closeness among inquiry picture yet dataset pictures are looked at the use of Euclidean separation. The general execution concerning proposed work is contrasted and territory about the triangle standardization technique for MPEG-7 CE Shape-1 Part-B dataset. From the empiric outcomes, we finish that standardization based concerning place about the area is additional precise.

Keywords: CBIR, Image retrieval, Equal Area Normalization, Feature extraction, Contour Normalization.

1. Introduction

Image data generated over the web is rising exponential system because final two decades. It is very a great deal crucial after develop fantastic image retrieval methods for dealing with and environment friendly processing. The common purpose about this technology methods are in conformity with fulfil user’s necessities and simplify the image retrieval task. Traditional Image search engines are based on keyword based search called

(2)

Text-Based Image Retrieval (TBIR). It needs extensive human comment and ethnical can edit boob within annotation so is different humans perceive the identical image in different affection and some image article are difficult after annotate exactly. These drawbacks are beat by introducing Content Based Image Retrieval (CBIR). The CBIR is the separate subject starts at the decenniad concerning 1960’s. CBIR automates photo retrieval by extracting photo features. The predominant strand about CBIR is after automatic image retrieval without any human intervention so is after redact wise machines for image retrieval.

The proposed action concerning CBIR using structure feature extraction, combines the similar fields concerning laptop science because of efficient retrieval on images: Image Processing (Image Segmentation, Image Signature Extraction) Database Technology (Query Processing, Storage, Retrieval concerning snap shots and picture features), and Data Mining (Similarity Measurement, Ranking over Results).

2.Literature Review

This portion gives a short survey about related action done within the field about content material based totally image retrieval. To date, there have been range regarding CBIR manufactory published; they mainly focus of characteristic extraction techniques.

2.1. Colour

The caricature personality recognition regulation raised by means of Jun Yu et al. [1,2,10] usage three features. The colour histogram, Hausdorff facet feature, then Skeleton characteristic are used after encode color, shape, then gesture facts on cartoon persona respectively.

2.2. Texture

Texture feature is a muscular partial descriptor so much helps of the retrieval process. One concerning the popular representations concerning earth feature is the co-occurrence cast who includes Gray level co-occurrence form (GLCM) or Color co-occurrence shape (CCM). These casts are primarily based over pixel orientation or inter-pixel distance. Tamura et al. [7,8] proposed a approach in accordance with expel field applications via the use of consequent visual texture homes specifically perceived lightness, uniformity, density, linearity, frequency, phase, randomness, fineness, smoothness, granularities, coarseness, contrast, directionality, likeliness, estimation and roughness. In CCM, acres applications are extracted out of every shade plane (R,G,B) in my opinion [8]. Curvelet captures ground services more effectively than lousy spectral purposes kind of Gobar yet Wavelet transforms.

2.3. Shape

Compared along other features, structure is considered the most hopeful for the identification on objects between an image. Before extracting the goal shape, suitable segmentation algorithm is applied. Segmentation methods can stay categorized in ternary wide categories: region-based, boundary-based, then pixel-based.

The motley education based structure retrieval the use of rot over difference measures with the aid of Chahooki et al. [9] integrates contour-based (Centroid and Farthest corner point) then region-based (Squared or Zernike moments) form retrieval methods. Because regarding 4 capabilities ancient the generated functions are massive size. The various education approach is used dimensionality discount on characteristic vectors among a non-linear manner.

2.4 Permanency Local Invariant Features

The disadvantage of low-level features such as colour, texture and form for image retrieval is 1) dropping a lot element statistics of the images, within suit of searching because of images as include the same goal or equal aspect with one-of-a-kind viewpoints 2) poverty about robustness according to occluded objects cognizance

(3)

and 3) dropping its pliability in imitation of light changes. The activity factor detectors yet descriptors are employed of many CBIR systems according to beat the above flaws. The Speed Up Robust Feature (SURF) is some regarding the almost yet famous hobby point detector or descriptor as has been posted via Bay et al. [3,4].

Challenges and Issues

There are couple types concerning CBIR specifically General CBIR and Application Specific CBIR. It is convenient in imitation of solve a distinctive lawsuit concerning CBIR (Medical, Biometric, Forensic, Industrial, Security, etc) than fake in imitation of remedy the overall CBIR hassle because like is nonetheless no universally frequent efficient yet significant image segmentation, function extraction, indexing yet retrieval strategies on hand for CBIR. Second difficulty is perplexity among looking for comparable images then searching for comparable objects after retrieve. Next issue is such is very challenging in imitation of become aware of and understand occluded objects out of the images[5]. The current CBIR structures lacking among query system mechanisms due to the fact input for CBIR systems is images, such is tough because of the stop person to structure photograph queries. Another one necessary difficulty is giant semantic hole between vile degree photograph features (CBIR truely works) yet excessive level semantic information present of the image (actually the person wants). So modern-day CBIR structures have not but had substantial impact concerning community appropriate after an incapability after deck bridge the semantic gap among computers and humans.

2.5 Toughness Problem Description

Durability A complete instruction has been made within the present CBIR structures into who entire structures are lacking of incomplete kind concerning person required functionalities. Among the old function extraction methods, structure is considered the most hopeful because of the identification on objects between an image.

Many structure feature extraction techniques raised in the literature. The Equal Area Normalization is a contour based range area structure characteristic extraction method is accessible into the current literature [1]. This technique divides the goal vicinity within equalize vicinity segments the usage of triangle vicinity formula. The proposed method replaces triangle region method including segment area approach because Equal Area Normalization, such performs properly because recognizing unrestricted form objects and normal shaped objects. The feature vector produced is some dimensional yet the Euclidean distance metric is old because of agreement measurement

3. Equal Area Normalization

The essential focus regarding proposed work is function extraction from MPEG-7 interior experiment CE-Shape-1 dataset [6] images and query image. The contour-based structure function extraction technique is used. The present work proposed with the aid of YANG et al [1, 2] applied an algorithm so much includes joining essential phases: Equidistant vertices normalization yet Equal Area Normalization (EAN). The EAN is no longer consequently a lot complicated in contrast in imitation of mean structure descriptor consequently such requires less computational resources. The proposed works do also identify objects concerning specific posses or sizes about still snap shots then video scenes. This work is at all useful because of object tracking and rule applications [10].

(4)

Figure 3.1. Schematic diagram of proposed CBIR system

The figure. 3.1 shows schematic design over proposed CBIR system, which carries joining fundamental sections: on-line process yet off-line process. The agreement metering arrest computes the strip into query and dataset images.The detailed algorithm given below describes operational concepts of EAN [11].

Algorithm:

Input: Query image; Output: Retrieval result 1. Input the query image.

2. Apply preprocessing techniques like image resize, zero padding, noise removal, holes removal.

3. Extract the contour. The contour is normalized by equidistant vertices.

3.1 Normalize the continuous contour into N equal distance points (eg) Assume N=64 using equidistant vertices(i.e., Pµ, Pµ+1…).

3.2 Find centroid G and move it to origin of the system.

3.3 Find Total Area of the contour ‘S’ by summing all triangular Area between three points (2 successive points on contour and centroid point).

4. Now the contour is normalized by EAN. Assume new points in EAN are Pt . 4.1 First divide the total area by N (eg) N=64. Spart=S/N

4.2 Select arbitrary starting point as the starting point used in equidistant vertices normalization (Pµ ). Compute the area of (Pµ, Pµ+1,G) and compare with Spart.

(5)

if (Area (Pµ, Pµ+1, G) == Spart)

Make the point Pµ+1 as Pt+1(a point in EAN) else

Seek the point Pt+1 on the segment (Pµ+1, Pµ+2)

4.3 Find the next adjacent area called (Pt+1, Pt+2, G) by mark Pt+1 as first vertex and repeat the above step to mark Pt+2.

4.4 Above two steps(4,2,4.3) are repeated to find all new points on the contour for EAN. The output is part area vector contains equal area segment values

5. The normalized equal area vector is obtained by area normalization and min-max normalization.

6. Compare the feature vectors using Euclidean distance formula.

7. Retrieve and display the images based on the relevance of input query image.

The proposed work of EAN involves the following modules:

1. Image preprocessing

2. Equi distant vertices normalization 3. Equal Area Normalization

A. Image preprocessing:

Text ter:

The image preprocessing step enhances makes the input image appropriate because of similarly steps. In that task image resize, nothing padding, clamor removal, holes elimination operations are applied.

In the fig. 3.2 shown, the input(butterfly-20.gif) image of quantity 419X402 pixels and resized image bulk is 264 X 254. permanency After preprocessing, the goal limitation pixels are extracted within supplement the usage of facet similar algorithm.

B. Equi distant vertices normalization

The following steps are performed in this module:

1. Starting point identification 2. Find total contour distance

3. Perform Equi distant vertices normalization 4. Find centroid of the contour

The Starting point on the contour is marked by two ways:

1. Farthest point on the contour from centroid.

2. Choose the first non-zero pixel value as starting point.

(6)

Input Query Image

Resized Image Zero Padded Images

Noise Free Image Holes Free Image

Fig. 3.2 Output of Image Preprocessing

Total contour distance or perimeter is obtained by summing the distance between adjacent boundary pixels beginning from the starting point using Euclidean distance [12]. The distance between two successive points may be 1 for orthogonal neighbour or 1.414 for diagonal neighbour. The formula to find total contour distance is,

 

kn xn xn yn yn Dt

Total_ 1 ( 1 )2 ( 1 )2

The equal distance pixels are computed by dividing Total contour distance by user defined value N. Now the normalized equal distance points on the contour is N. The formula is,

N Dt Total Dt

Equal _

_ 

Centroid is the geometric centre of a plane object. It is the point in a plane area such that the moment of the area, about any axis, through that point is zero. The Centroid of an image is rotationally invariant. The Centroid was calculated using second order moments formula given below:

00 10

m xm

and 00

01

m ym

Centroid (G) = { x , y }

Where, m00

-Zeroth order moment and m10 , m01

-First order moments.

C. Equal Area Normalization

The major activities of this module is to, 1. Find total contour area

2. Extract features(EAN)



 

 

2 2 1

sect 2 2

) 1 Sector(S of

Area r r

(7)

The first step in this module is to find the total area of the object. The total area of the object is computed by taking summation of all triangle areas with points Pµ, Pµ+1, G (i.e. Centroid). The Total Area and Part Area are computed using two methods:[13]

1. Triangle Area Method

2. Segment (Sector) Area Method

In triangle area method, the distance between two points and represented in following figure, the formula for area of the triangle is also given below.

G(x,y)

For easy computation, the centroid is transformed to origin(0,0) and then calculating the Total Area of the Object using Summation of area of all triangles. By simplifying,

Where is the area of the triangle vertices and G.

In Segment Area method, the area of the sector between three points centroid G and two points on the contour A, and B is represented in the following figure.

The area of the sector formula is, Where,

ᶿ - Angle between the points A and B measured from centroid G, r1,r2- Distance(radius) from G to A and B.

(8)

Thus the total area of the object can be found by the summation of area of all sectors.

Ssect Area(S)

Total

Up to this point, there are two ways the total contour area value is found. Now divide the total area by N as mentioned in algorithm step 4.1.

Spart=S/N

From the algorithm steps 4.2 to 4.4 generates two different part area vectors (EAN1, EAN2) for each image using triangle area method and segment area method respectively:

EAN1={A1part0, A1part1, A1part2, …. A1partN-2, ApartN-1} and EAN2={A2part0, A2part1, A2part2, …. A2partN-2, A2partN-1} Where

A1part – Triangle based part area calculated between centroid and two points on the contour.

A2part – Sector based part area calculated between centroid and two points on the contour.

4. Experimental Result

The proposition on shape descriptor is tested on MPEG-7 CE Shape-1 Part-B dataset, which contains 1400 GIF format binary images with 70 classes each class with 20 different images. When a query image comes in, it is segmented and its features are extracted and stored. The features of the query image are compared with features of all the images in the database using Euclidean Distance.

The output of EAN is part area vectors contains N equal part area values. These values are normalized by two methods:

(1) Area Normalization (2) Min-Max Normalization

These normalized values are narrow down to certain range; and it is useful for easier analysis of data.

A. Area Normalization

In area normalization we get factor values by dividing part area values (Apart) by total area (A).







A A

A

v' Apart0 part1 part2 ApartN-1

,....

,A ,A A

Where,

v - Normalized Equal Area Vector '

Apart- Equal Area Vector A- Total Object Contour Area Min-Max Normalization

(9)

The part area values (Apart) in the feature vector is mapped in new range [New_MinA,New_MaxA] by using the following formulae

A A

A

A A

A New Max New Min New Min

Min Max

Min

v' v _  _  _

 

Where

Min - Minimum value present in the part area vector A

Max - Maximum value present in the part area vector A

v- Input part area values

v - Normalized Equal Area Vector. '

The similarity measurement is computed in EAN1{Query Image and Dataset Image} and EAN2{Query Image and Dataset Image} using Euclidean distance metric. The Euclidean distance formula is

2

2 ( )

) (

) ,

( Query Dataset Query Dataset

E Query Dataset X X Y Y

D    

Where

(XQuery, YQuery) – Normalized part area value of Query Image, (XDataset

,YDataset

) – Normalized part area value of Dataset Image.

The fig. 4.1 shows execution of query and dataset image, its comparison graph and distance values by area normalization and min-max normalization. From the graph and distance values it is observed that segment area method performs better than the triangle area method.

Query Image Dataset Image Comparison Graph Distance Values Area Normalization

Min-Max Normalization Triangl

e Area

Segment Area

Triangle Area

Segment Area

bell-3.gif beetle-20.gif 0 10 20 30 40 50 60 70 0.0145

0.015 0.0155 0.016 0.0165 0.017 0.0175

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0107 0.0101 1.1127 1.1113

(10)

bell-3.gif bell-20.gif 0.01480 10 20 30 40 50 60 70

0.015 0.0152 0.0154 0.0156 0.0158 0.016 0.0162 0.0164 0.0166

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0089 0.0089 1.0105 1.0240

bell-3.gif bell-3.gif 0.01480 10 20 30 40 50 60 70 0.015

0.0152 0.0154 0.0156 0.0158 0.016

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0000 0.0000 0.0000 0.0000

cattle-6.gif cattle-12.gif 0.01560 10 20 30 40 50 60 70

0.0157 0.0157 0.0158 0.0158 0.0159 0.0159 0.0159 0.016 0.016

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0027 0.0025 0.9544 0.9650

cattle-12.gif children-1.gif 0.0140 10 20 30 40 50 60 70

0.0145 0.015 0.0155 0.016 0.0165 0.017 0.0175 0.018

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0088 0.0087 1.0174 1.1031

cattle-2.gif device7-1.gif 0 10 20 30 40 50 60 70

0.0155 0.016 0.0165 0.017 0.0175 0.018 0.0185

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0193 0.0194 0.8690 1.1741

children-1.gif children-11.gif 0.0140 10 20 30 40 50 60 70

0.0145 0.015 0.0155 0.016 0.0165 0.017 0.0175

0.018 Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0095 0.0086 0.8051 0.7505

children-1.gif chopper-12.gif 0.0140 10 20 30 40 50 60 70

0.0145 0.015 0.0155 0.016 0.0165 0.017 0.0175

0.018 Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0103 0.0103 0.9261 0.9319

tree-13.gif tree-8.gif 0.01450 10 20 30 40 50 60 70

0.015 0.0155 0.016 0.0165 0.017 0.0175 0.018

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0074 0.0074 1.0188 0.9563

(11)

tree-13.gif apple-6.gif 0.01540 10 20 30 40 50 60 70

0.0155 0.0156 0.0157 0.0158 0.0159 0.016

Area Normalization Using Triangular Area and Segment Area

Number of Points

Normalized Difference Value

0.0052 0.0052 1.0274 1.0206

Fig 4.1 Sample experimental results with distance values

5. Conclusion

We have shown that, a contour based structure descriptor based on Equal Area Normalization provides a better solution because of objective awareness troubles and content material based image retrieval systems design. The overall performance over the proposed approach is compared with honour according to two techniques as partition destination location into equal bulk segments are triangle place method, and area over the phase method. The empirical results indicate that, into near cases, the proposed rule along region of the phase method execute drastically outperforms properly about extraordinary object shapes among MPEG-7 CE Shape-1 Part-B dataset. The characteristic route about it labor is in imitation of understand multiple objects from cluttered frequent spectacle images.

References

1. YANG Mingqiang, KPALMA Kidiyo, and RONSIN Joseph “Affine Invariance Contour Descriptor Based on the Equal Area Normalization”, IAENG International Journal of Applied Mathematics, May 2007.

2. YANG Mingqiang, KPALMA Kidiyo, and RONSIN Joseph “Shape-based Invariant Feature Extraction for Object Recognition”, Advances in reasoning-based image processing, analysis and intelligent systems: Conventional and intelligent paradigms, 2012.

3. Remco C. Veltkamp, and Mirela Tanase, “Content-Based Image Retrieval Systems: A Survey”, Technical Report UU- CS,October 28, 2002.

4. John Eakins, “Content-based Image Retrieval” The JISC Technology Applications Programme Report 39, October 1999.

5. Sadegh Abbasi, Farzin Mokhtarian, Josef Kittler “Curvature Scale Space image in Shape Similarity Retrieval”, Multimedia Systems-Springer Varlag 1999, Page-467-476.

6. Swati V. Sakhare, and Vrushali G. Nasre, “Design of Feature Extraction in Content Based Image Retrieval (CBIR) using Color and Texture”, International Journal of Computer Science & Informatics, Volume-I, Issue-II, 2011.

7. P. Gangadhara Reddy, “Extraction of Image Features for an Effective CBIR System”, IEEE, 978-1-4244-9182-7/10, pages 138-142, 2010.

8. K.Velmurugan, and Dr.S. Santhosh Baboo, “Content-Based Image Retrieval using SURF and Colour Moments”, Global Journal of Computer Science and Technology Volume 11 Issue 10, 2011.

9. Ahamed, B. B., & Ramkumar, T. (2015). Deduce User Search Progression with Feedback Session. Advances in Systems Science and Applications, 15(4), 366-383.

10. Ahamed, B. B., & Ramkumar, T. (2018). Proficient Information Method for Inconsistency Detection in Multiple Data Sources.

(12)

11. Ahamed, B. B., Ramkumar, T., & Hariharan, S. (2014, December). Data integration progression in large data source using mapping affinity. In 2014 7th International Conference on Advanced Software Engineering and Its Applications (pp. 16-21). IEEE.

12. Ahamed, B. B., & Hariharan, S. (2011). A survey on distributed data mining process via grid. International Journal of Database Theory and Application, 4(3), 77-90.

13. Ahamed, B. B., & Hariharan, S. (2012). Implementation of Network Level Security Process through Stepping Stones by Watermarking Methodology. International Journal of Future Generation Communication and Networking, 5(4), 123-130

References

Related documents

(Hare's account does not fit.. into this pattern, but will be analysed in detail later in this chapter.) Another strand of the tradition tries to avoid the obvious problems

ANNEX ACE Life In Two Stamf 281 Tresse Stamford, C Main Telep NAIC Com JURISDICT Alabama Alaska Arizona Arkansas California Colorado Connecticu Delaware District of C Florida

Maintaining a rigorous standard for enhanced damages would limit the impact of this type of activity because innovators could navigate these scenarios more confidently. They would

The areas of responsibility of the Coordination Committee for Hospitals and University comprise collaboration on research, talent development, education and knowledge

With this back- ground, this study was conducted with the following objectives: To show that PCT is an accurate marker for differentiating OM and SA from viral infections and

and K.. This is accounted for by the presence of adsorbed on the surface of Pt black saturated with separate Pt oxides in the mixture of T i and Pt H, showed that two

The above results indicate that the five heavy metals have low accumulation in most of the propagules of 10 mangrove species and are at safe levels for the

Effects of development of the steady boundary layer flow, heat transfer and nanoparticle volume fraction over a stretching surface in a nanofluid are studied for