MSR-Bing Image Retrieval Challenge ,written by Win

25
Text Image Retrieval Challenge -Enhance relationships between query and image InstructorMeiChen Yeh ChenLin Yu, ChiungWei Hsu VIPLAB

description

For 2014 MSR-Bing Image Retrieval Challenge written by Win

Transcript of MSR-Bing Image Retrieval Challenge ,written by Win

Page 1: MSR-Bing Image Retrieval Challenge ,written by Win

Text

Image Retrieval Challenge -Enhance relationships between query and image

Instructor: MeiChen Yeh ChenLin Yu, ChiungWei Hsu

VIPLAB

Page 2: MSR-Bing Image Retrieval Challenge ,written by Win

Outline

1. Proposed method

2. Evaluation Metric

3. Experiment Result

4. Finding and Difficulty

5. Demo

6. Conclusion

7. Future work

Page 3: MSR-Bing Image Retrieval Challenge ,written by Win

Proposed Method

Page 4: MSR-Bing Image Retrieval Challenge ,written by Win

Query

Natural Language Processing

Tokenization

POSt

QE by WordNet QE by WikipediaWordNet Wikipedia

Click_count ranking Top candidatesUser Clicklog from

MSR dataset

Page 5: MSR-Bing Image Retrieval Challenge ,written by Win

Apple apple apples

an apple ….

Page 6: MSR-Bing Image Retrieval Challenge ,written by Win

Query Processing

1. Stop word and removal

2. Tokenization

3. Stemming and Lemmatization

4. Part-of-speech Tagging

5. Wiki-suggestion (Misspelled words)

6. Expansion (wordnet and wikipeia)

Page 7: MSR-Bing Image Retrieval Challenge ,written by Win

Apple apple apples

an apple ….

Ranking Tablelog candidate count image

apple 1890 QYQtQsx9lH1KwA

apple 503 QJ4gfSPJYhbw0A

… … …

apple mac 490 PvfGna70qGiBIA

Page 8: MSR-Bing Image Retrieval Challenge ,written by Win

Click-count Ranking

MSR dataset provide real world data for user query log.

With this, generated homemade searching table by“Click-count”.

“Max click count rule”

Log data 1,000,000 (only 1/20)

We can make sure that candidate pictures are most popular.

Page 9: MSR-Bing Image Retrieval Challenge ,written by Win

Apple apple apples

an apple ….

Ranking Tablelog candidate count image

apple 1890 QYQtQsx9lH1KwA

apple 503 QJ4gfSPJYhbw0A

… … …

apple mac 490 PvfGna70qGiBIA

Page 10: MSR-Bing Image Retrieval Challenge ,written by Win

Evaluation Metric

Page 11: MSR-Bing Image Retrieval Challenge ,written by Win

MSR vs DIY Method

!

!

[rel]={Excellent=3,Good=2,Bad=0}

X

Page 12: MSR-Bing Image Retrieval Challenge ,written by Win

Experiment Result

Page 13: MSR-Bing Image Retrieval Challenge ,written by Win

Prepare and WorkOff-line:

NLTK to process user query log

Build Ranking table (1,000,000)

Include image(base64) to Database(800,000)

On-line:

NLTK to process query input

Query expansion by word net and wikipedia

Large-scale database query processing

Page 14: MSR-Bing Image Retrieval Challenge ,written by Win

Single unit-query'president','frank','mars','chinese','taiwan','dargon','crash','bird','France','Eiffel','president','tony','frank','mars','chinese','taiwan','London','Mexican','ydney',

'google','yahoo','jessica','microsoft','amazon','windows','apple','line','linux','android',

'world','iphone','bacteria','cat','basketball','dog','micky','tom','jerry','christmas','table',

Test : 32 queries Acc:87.5 %

Page 15: MSR-Bing Image Retrieval Challenge ,written by Win

Compound word-querybook store, picture frame, the lost and bewildered tourist, ice cream, cell phone, apple pie, a story as old as time, a cool wet afternoon, many cases of infectious disease

swimming pool, the senlie old man,pencil box , long and winding road, tiddy bear , hot dog, jennifer love hewitt, some cookie shaped like stars

hello kitty coloring page, kelly osbourne drinking, micky mouse, a wet amd stinky dog

Test : 20 queries Acc:42.28 %

Page 16: MSR-Bing Image Retrieval Challenge ,written by Win

Finding and Difficulty

Page 17: MSR-Bing Image Retrieval Challenge ,written by Win

Spelling correctly can improve retrieval accuracy.

Query expansion can find more related images

!

A ambiguous query can be difficult to used.

The gap exists between users and result images, because the word is polysemic.

The user query still has a semantic problem.

Finding

Page 18: MSR-Bing Image Retrieval Challenge ,written by Win

In a compound word query, the relationship

between previous and next word is very

important.

Query semantic is still a challenge.

Large-scale data processing is a big problem.

How to speed up search performance?

Difficulty

Page 19: MSR-Bing Image Retrieval Challenge ,written by Win

Demo

Page 20: MSR-Bing Image Retrieval Challenge ,written by Win

Conclusion

Page 21: MSR-Bing Image Retrieval Challenge ,written by Win

Enhance relationships between query and image

Page 22: MSR-Bing Image Retrieval Challenge ,written by Win

Find relationships between query and image

Page 23: MSR-Bing Image Retrieval Challenge ,written by Win

Future Work

Page 24: MSR-Bing Image Retrieval Challenge ,written by Win

Query

Natural Language Processing

Tokenization

POSt

QE by WordNet QE by WikipediaWordNet Wikipedia

Click_count ranking Top candidates

Named Entity Recognition

User Clicklog from MSR dataset

Enhance

Page 25: MSR-Bing Image Retrieval Challenge ,written by Win

–ChenLin Yu, ChiungWei Hsu

“Thank you”