+ speed up texture classification in clothing retrieval system 電機三 吳瑋凌

17
+ Speed Up Texture Classification in Clothing Retrieval System 電電電 電電電

Upload: emmeline-alexander

Post on 06-Jan-2018

233 views

Category:

Documents


5 download

DESCRIPTION

+ What’s the data in images? HOG LBP Color Histogram Color Moment Skin

TRANSCRIPT

Page 1: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+

Speed Up Texture Classification

in Clothing Retrieval System

電機三 吳瑋凌

Page 2: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+What is clothing retrieval system?When a clothing image comes,

we’d like to find the same pictures or the similar ones.

Page 3: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+What’s the data in images?

HOGLBPColor HistogramColor MomentSkin

Page 4: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+Motivation

Large scale computing efficiency

=> want to find a way to speed up the system

Page 5: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+MethodFind some conditions to classify

imagesOnly need to compute features with

those in the same class

Page 6: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+Experiment

Use HOG to classify different texturesDefine 5 groups and use 25 training

pictures for each

Page 7: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+HOG(Histogram of oriented gradients )Detect the edge of items Judge clothing texture 3*3*9=81 dimensions

z

cellblock

Page 8: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+How to find conditions

threshold

Find a threshold making the precision high enough after splitting

Page 9: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+How to find conditions

Don’t define threshold

Threshold=(max1+min2)/2

Page 10: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+ExperimentOne dimension at most 10 threshold

valuesFind the one having the max precision

in all dimensions Set Paim

<threshold >threshold

Page 11: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+ClasstreePaim=0.

7

Page 12: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+Results5 pictures for each group(total=25

pictures)Calculate precision and recall for each

group

Page 13: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+DiscussionThe relationship of P,R and PaimThe higher Paim, the higher P and R

Paim

Page 14: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+DiscussionUse F1 to evaluate F1=2*P*R/(P+R)

Paim

Page 15: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+DiscussionThe relationship of training time and

PaimThe higher Paim, the longer training

time

Page 16: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+Conclusion

We can find some more important entry to classify some groups

Then we can only compute features of query image with those in the same groups

Page 17: + Speed Up Texture Classification in Clothing Retrieval System 電機三 吳瑋凌

+Thanks for your listening :)