Upcoming SlideShare
×

# Feature Extraction

5,524 views

Published on

3 Likes
Statistics
Notes
• Full Name
Comment goes here.

Are you sure you want to Yes No
• nice work

Are you sure you want to  Yes  No
• thinks

Are you sure you want to  Yes  No
• merci beaucoup

Are you sure you want to  Yes  No
• it is very useful

Are you sure you want to  Yes  No
Views
Total views
5,524
On SlideShare
0
From Embeds
0
Number of Embeds
6
Actions
Shares
0
265
4
Likes
3
Embeds 0
No embeds

No notes for slide

### Feature Extraction

1. 1. Two Feature Extraction Methods Lian, Xiaochen skylian1985@163.com Department of Computer Science Shanghai Jiao Tong University July 13, 2007 Lian, Xiaochen Two Feature Extraction Methods
2. 2. Attention Based Method Statistics Based Method Outline 1 Attention Based Method Why Attention? Model of Attention Application in Face Recognition 2 Statistics Based Method Basic Idea Feature Selection Process Lian, Xiaochen Two Feature Extraction Methods
3. 3. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Outline 1 Attention Based Method Why Attention? Model of Attention Application in Face Recognition 2 Statistics Based Method Basic Idea Feature Selection Process Lian, Xiaochen Two Feature Extraction Methods
4. 4. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Why Attention? When recognizing a person, we compare the face with those stored in memory. We always can not remember all the details of a face. It is the conspicuous parts that impress themselves on us. Lian, Xiaochen Two Feature Extraction Methods
5. 5. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Model of Attention How do human vision system ﬁnd salient regions in a scene? Koch and Ullman[?] proposed a biologically plausible architecture. Figure: General architecture ofExtraction Methods Lian, Xiaochen Two Feature the model
6. 6. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Channels The original image is decomposed into three channels. Intensity I : Consider the brightness of a pixel, which is obtained as I = (r + g + b)/3. Color: Red-Green color and Blue-Yellow opponencies. r−g RG = max(r, g, b) b − min(r, g) BY = max(r, g, b) Orientation: Four orientation channels correspond to gabor ﬁlters oriented at 0, 45, 90, and 135 degrees. This representation is able to capture the critical distinctions in orientation. Lian, Xiaochen Two Feature Extraction Methods
7. 7. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Channels Figure: Channels Lian, Xiaochen Two Feature Extraction Methods
8. 8. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Image Pyramid The Gaussian pyramid is created to a depth of nine levels, with level 0 having a scale of 1 : 1 (the original input image) and level 8 being 1 : 256. This is done by ﬁltering the images with gaussian ﬁlter and then resize it. We use gaussian ﬁlter to eliminate noise, and the resizing is for biological purpose. There are seven pyramids, one for intensity MI , two for color MRG and MBY , and four for orientation Mθ (θ ∈ {0, 45, 90, 135}). Lian, Xiaochen Two Feature Extraction Methods
9. 9. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Image Pyramid Figure: Image pyramid Lian, Xiaochen Two Feature Extraction Methods
10. 10. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Center Surround Difference It is a cross-scale difference between two images, denoted by “ ”: expanding the smaller image into the larger one by interpolation, then followed by pixel-pixel substraction. Figure: Center Surround Difference Lian, Xiaochen Two Feature Extraction Methods
11. 11. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Normalization Figure: Normalization effect Lian, Xiaochen Two Feature Extraction Methods
12. 12. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Normalization Difference-of-Gaussians ﬁlter is usually used to detect blob. c2 −(x2 +y2 )/(2σ2 ) c2 e−(x +y )/(2σinh ) 2 2 2 DoG(x, y) = ex 2 e ex − inh 2 2πσex 2πσinh Lian, Xiaochen Two Feature Extraction Methods
13. 13. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Saliency Map Combine the images from all the channels linearly. Figure: Saliency Map Lian, Xiaochen Two Feature Extraction Methods
14. 14. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Figure: Faces and the corresponding saliency Map(from ORL face database) Lian, Xiaochen Two Feature Extraction Methods
15. 15. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Experiment Result Figure: Error rate Figure: Rank error rate Lian, Xiaochen Two Feature Extraction Methods
16. 16. Why Attention? Attention Based Method Model of Attention Statistics Based Method Application in Face Recognition Lots of Problems! How to do recognition? Different people have different sets of features. Simply applying Euclid Distance yields bad performance: the error rate is high for a 40-person database. The performance suffers pose and expression severely. Lian, Xiaochen Two Feature Extraction Methods
17. 17. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Outline 1 Attention Based Method Why Attention? Model of Attention Application in Face Recognition 2 Statistics Based Method Basic Idea Feature Selection Process Lian, Xiaochen Two Feature Extraction Methods
18. 18. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Basic Idea Suppose S = {x1 , x2 , · · · , xn } be n features for the collected data. The objective of feature selection is to ﬁnd a subset Sd = {xi1 , xi2 , · · · , xid }, which sufﬁce to represent the original data. The performance of Sd can be evaluated by the percentage of the variation in xi that can be accounted for by the elements by Sd . If that percentage is large enough, Sd can then be the ﬁnal choice; otherwise, new signiﬁcant variables need to be added into Sd . Lian, Xiaochen Two Feature Extraction Methods
19. 19. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Feature Similarity Measure The squared-correlation coefﬁcient between two random vectors x and y is (xt y)2 sc(x, y) = . (xt x)(yt y) This measure has the following properties: 0 ≤ |sc(x, y)| ≤ 1. |sc(x, y)| if and only if x and y are linearly related. The measure is invariant to scaling and translation. The measure is sensitive to rotation. Lian, Xiaochen Two Feature Extraction Methods
20. 20. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Step-By-Step Selection At the ﬁrst step, let i=1 sc(xi , xj ) n Cj,1 = , n i1 = arg max {Cj,1 }. 1≤j≤n Select xi1 as the ﬁrst signiﬁcant variable. Lian, Xiaochen Two Feature Extraction Methods
21. 21. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Step-By-Step Selection Assume the ﬁrst m − 1 most signiﬁcant variables, z1 , · · · , zm−1 , has been chosen. The m-th signiﬁcant feature zm will be chosen in such a manner: The subset Sm−1 + {zm } should be the most representative subset compared with any other subsets formed by adding a candidate feature to Sm−1 . Let αj ∈ S − Sm−1 and i=1 sc(xi , αj ) n Cj,m = , n im = arg max {Cj,m }. 1≤j≤n The m-th signiﬁcant feature can then be xim . Lian, Xiaochen Two Feature Extraction Methods
22. 22. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Lian, Xiaochen Two Feature Extraction Methods
23. 23. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Some Discussion The squared-correlation coefﬁcient is used to measure the linear correlation between variables. Need new method for nonlinear relationships. The greedy search process do not assure the optimal selection. The complexity is O(n2 N), where n is the number of features, and N is the number of samples. When n become large, the algorithm will be inefﬁcient. Lian, Xiaochen Two Feature Extraction Methods
24. 24. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process C. Koch, and S. Ullman, “Shifts in Selective Visual Attention: Towards the Underlying Neural Circuitry,” Human Neurobiology, vol. 4, pp. 219-227, 1985. pp. 89-102, 1977. Laurent Itti, and Christof Koch, “A saliency-based search mechanism for overt and covert shitfs of visual attention,” Vision Research, 40(2000). Dirk Walther, and Christof Koch, “Modeling attention to salient proto-objects,” Neural Networks, 19(2006). Laurent Itti, Christof Koch, and Ernst Niebur, “A Model of Saliency-Based Visual Attention for Rapid Scene Analysis,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 20, No. 11, Nov. 1998. Lian, Xiaochen Two Feature Extraction Methods
25. 25. Attention Based Method Basic Idea Statistics Based Method Feature Selection Process Hua-Liang, and Stephen A. Billings, “Feature Subset Selection and Ranking for Data Dimensionality Reduction,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 29, no.1, Jan. 2007. Pabitra Mitra, C.A. Mrthy, and Sankar K. Pal “Unsupervised Feature Selection Using Feature Similarity,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 24, No. 3, Mar. 2002. Lian, Xiaochen Two Feature Extraction Methods