발표자료

Report
Image Auto Tagging 기술 개발
2014. 3. 5
삼성전자 종합기술원
Device & System 연구센터
S/W Solution Lab.
- 0/2 -
Image Auto Tagging 기술 개발
• Goal
– 상품 분류를 위한 Image Automatic Tagging 기술을 개발
• Approach
– 상품 Image와 Tag 정보를 (이름/Description) Sample Image, Crawling 등을 통해 확보
– Tag 정보와 Image 간의 연관성 Model을 확보 ( Deep Learning, Text-Image 연관분석 등)
– 확보된 Model의 유효성 검증 (확보된 Sample 영상 Test 및 새로운 Web Crawling 영상 Test)
• 필요 기술
– Image/Text Data Mining
• Image feature extraction, data clustering , similarity search
– Machine Learning
• Multi-class classification ( deep brief network)
• 개발 기간/인원/Contact
– 12개월 (4인)
– 업체 담당자: 삼성전자 종합기술원 박진만 전문연구원 (전화: 010-7327-9191)
- 1/2 -
참고 : Image Classification (ILSVRC2013)
• Large Scale Visual Recognition Challenge 2013
Number of object classes
Training
Validation
Testing
200
Num images
395909
Num objects
345854
Num images
20121
Num objects
55502
Num images
40152
Num objects
---
Team name
Comment
Error
Clarifai
Multiple models trained on the original data plus an additional model
trained on 5000 categories.
0.11197
NUS
adaptive non-parametric rectification of all outputs from CNNs and
refined PASCAL VOC12 winning solution, with further retraining on
the validation set.
0.12953
ZF
5 models (4 different architectrues) trained on original data.
0.13511
Andrew Howard
This is an ensemble of convolutional neural networks combining
multiple transformations for training and testing and models
operating at different resolutions.
0.13555
* : http://www.image-net.org/challenges/LSVRC/2013/
- 2/2 -

similar documents