31
Splitting and Merging Approach Splitting and Merging Approach for Image Indexing and Retrieval for Image Indexing and Retrieval in DC Domain in DC Domain By :S. Irianto Supervisor Prof. Jianmin Jiang Department of EIMC School of Informatics

Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

  • Upload
    dacian

  • View
    33

  • Download
    2

Embed Size (px)

DESCRIPTION

Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain. By :S. Irianto. Supervisor Prof. Jianmin Jiang. Department of EIMC School of Informatics. MOTIVATION. a growing on demand for image database applications more people have accessed to large databases - PowerPoint PPT Presentation

Citation preview

Page 1: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Splitting and Merging Approach for Image Splitting and Merging Approach for Image

Indexing and Retrieval in DC DomainIndexing and Retrieval in DC Domain By :S. Irianto

SupervisorProf. Jianmin Jiang

Department of EIMCSchool of Informatics

Page 2: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

MOTIVATION MOTIVATION a growing on demand for

image database applicationsmore people have accessed to

large databases the research community to

develop methods to archive, query and retrieve this database based on their content.

Page 3: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

THE AIMSTHE AIMS

TO PRESENT AN EFFECTIVE IMAGE RETRIEVAL ON NON-DCT IMAGES

TO OFFER AN ALTERNATIVE METHOD OF IMAGE RETRIEVAL

TO INTRODUCE SEGMENTATION BASED FOR IMAGE RTREIVAL

Page 4: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

RELATED WORKSRELATED WORKS

VIRS (Gupta,1997)QBIC - wwwqbic.almaden.ibm.com

Photobook- vismod.media.mit.edu

Virage – www.virage.comCandid -

public.lanl.gov/kelly/CANDID/method.shtml

Page 5: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

RELATED WORKSRELATED WORKS …… ……

RetrievalWare – www.excalib.comVisualSEEK and WebSEEK – WWW.columbia.edu

Netra – vivaldi.ece.ecsb.edu/Netra/MARS – jadzia.ifp.uiuc.edu:8000

Page 6: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

DC IMAGE REPRESENTATIONDC IMAGE REPRESENTATION

Instead of working on the full image very time consuming and complex.

we extract only one value on every block of the jpeg images.

7

0

7

0

),(8

1 ),(

i j

jcirfcrDC

Page 7: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

DC IMAGE ………..DC IMAGE ………..

DC image extracted basically consist of average of other 63 pixels

Page 8: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC DC DC DC DC DC DC DC

DC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

DC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

DC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC AC

AC AC AC AC AC AC AC ACDC image

>>3 + 128

DC IMAGE …….

8 X 8 DCT image

Page 9: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

RGB Grayscale DC image Segmented

DC AND SEGMENTED IMAGE - EXAMPLES

Page 10: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

DC vs Segmented imagesDC vs Segmented images

DC image Segmented image

Page 11: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

DC vs …………………..DC vs …………………..

DC image Segmented image

Page 12: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

SPLITTING AND MERGINGSPLITTING AND MERGING(adopted from Dubuisson1993)(adopted from Dubuisson1993)

First, we must split the image, start by considering the entire image as one region

If the entire region is coherent or has sufficient similarity, leave it unmodified

Page 13: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

SPLITTING ………….SPLITTING ………….

If the region is not sufficiently coherent, split it into four quadrants; and recursively apply these steps to each new region.

Page 14: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

ALGORITHMALGORITHMsplit phase: create an (empty) tree T with regions in its nodescreate a node for T (which is root and the only leave)put X in this node for all leaves L of T take region RL from L if RL is sufficiently homogenous then return RL in L if RL is not sufficiently homogenous then separate RL into four sub-regions R1

L,

R2L, R

3L, R

4L, of equal size

create four new leaves L1; L2; L3; L4 as sons of L put Ri L in L1; 1 ≤ i ≤ 4

Page 15: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

ALGORITHM ……………..ALGORITHM ……………..merge phase:

create an (empty) set S of (empty) trees for all leaves L of T add L as a new tree to S (with L as root and only leave) for any root L in S for any root L’ ≠ L in S if the segments RL in L and RL’ in L’ are neighboured and similar then create a new root L in S with sons L and L’

put the segment RL Ư RL’ into L’

Page 16: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

I1 I2

I3 I4

DC image(I)

First split

SPLITTING AND MERGING PROCESS

Page 17: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

I1 I2

I3 I41 I42

I43 I44

Second split

I1 I2

I3 I41 I42

I43 merge

SPLITTING AND MERGING PROCESS

Page 18: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

IMAGE QUERYIMAGE QUERY

user make a query an image, once a query is specified,

we score each image in the database on how closely it satisfies the query

Page 19: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

IMAGE QUERY …….IMAGE QUERY …….

The score for each atomic query (DC -image) is calculated by using Euclidean distance transform

Page 20: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

RESULTS ANALYSISRESULTS ANALYSIS

Constraint1,000 images consist of seven classes consists of “motorbike”, “building”, “car”, “cat”, “flower”, “mountain”, and “sky”.

images of size 2 to 10 K

Page 21: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

RESULTS ………..RESULTS ………..

(DC image) -our propose methodHighest precision at 0.89 for MOUNTAINLowest precision at 0.22 for CAT

DCT imagehighest precision at 0.65 for MOUNTAINLowest precision at 0.45 for CAT

Page 22: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

RESULTS ………..RESULTS ………..

Average precisionfor DC image : 0.59for DCT image : 0.52

Page 23: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Precision and recallPrecision and recall

Image retrieval effectiveness

0

0.2

0.4

0.6

0.8

1

0 0.02 0.04 0.06 0.08 0.1

Recall

Pre

cis

ion SaM image

RGB image

Page 24: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Precision and recall…………...Precision and recall…………...Image retrieval effectiveness

mountain

cat

sky

building

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.02 0.04 0.06 0.08 0.1

Recall

Pre

cis

ion SaM image

RGB image

Page 25: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Number of regions vs precisionNumber of regions vs precision

Precision and its number of regions

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 50 100 150 200 250 300 350 400

Number of regions

Prec

isio

n

precision

Linear (precision)

Page 26: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Number of regions …………….Number of regions …………….

Precision and its number of regions

mountain

cat

sky

building

bike

carflow er

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

0 50 100 150 200 250 300 350 400

Number of regions

Prec

isio

n

precision

Linear (precision)

Page 27: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

CONCLUDING REMARKCONCLUDING REMARKSaM approach can be used as an

alternative technique to improve the effectiveness of image retrieval, particularly for non full DCT based images

The evidence seems indicate that the split-merge on DC image approach demonstrates somewhat higher on precision than RGB approach as existing technique, even though, the precision is not significantly different

Page 28: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

FUTURE WORKSFUTURE WORKS

WE ARE SETTING UP SEGMENTATION ON DC IMAGES DIRECTLY FROM RGB IMAGE BY USING REGION GROWING AND SPLIT-MERGE SEGMENTATION

Page 29: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Thank you for your patientWish you enjoy

Page 30: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

NotesNotes

Homogenity measured by

Similarity (merge) measured by

Page 31: Splitting and Merging Approach for Image Indexing and Retrieval in DC Domain

Notes ……….Notes ……….

Homogenity measured by

(10) Threshold 2

Merge measured by

(10) Threshold - 21