Post on 31-Dec-2015
讲解人 : 崔 振2010.9.17
Supervised Translation-Supervised Translation-InvariantInvariant
Sparse CodingSparse Coding
[Jianchao Yang, Kai Yu, Thomas Huang]
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
Jianchao Yang
jyang29 @ifp.uiuc.edu
Image Formation & Processsing Group (IFP), University of Illinois at Urbana-Champaign (UIUC)
Ph.D. Candidate (06-Present, ECE, UIUC) ; Ph.D. Adviser: Prof. Thomas S. Huang
B.Eng (02-06, EEIS, USTC)
Publication(第一作者) CVPR : 4篇, 2 篇 oral TIP : 2篇 ECCV10 , 1篇 ICIP,1篇
Homepage: http://www.ifp.illinois.edu/~jyang29/
Kai Yu
Machine Learning researcher and the Head of Media Analytics Department at NEC Laboratories America. Inc..
Ph.D. Computer Science, University of Munich,Germany, January 2001 – July 2004.
B.Sc and M.Sc, Nanjing University.
Research Interests Areas: machine learning, data mining, information
retrieval, computer vision CVPR(4),ECCV(4+),ICML(8+),NIPS(10+),…
http://www.dbs.informatik.uni-muenchen.de/~yu_k/
Thomas Huang
Beckman Institute Image Formation and Processing and Artificial Intelligence groups.
William L. Everitt Distinguished Professor in the U of I Department of Electrical and Computer Engineering and the Coordinated Science Lab (CSL);
Sc.D. from MIT in 1963
computer vision, image compression and enhancement, pattern recognition, and multimodal signal processing.
http://www.beckman.illinois.edu/directory/t-huang1
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
文章信息
文章出处 CVPR10 ( oral)
相关文章 Yang et al. Linear spatial pyramid matching using
sparse coding for image classification. CVPR’09.
Abstract
In this paper, we propose a novel supervised hierarchical sparse coding model based on local image descriptors for classification tasks. The supervised dictionary training is performed via back-projection, by minimizing the training error of classifying the image level features, which are extracted by max pooling over the sparse codes within a spatial pyramid. Such a max pooling procedure across multiple spatial scales offer the model translation invariant properties, similar to the Convolutional Neural Network (CNN). Experiments show that our supervised dictionary improves the performance of the proposed model significantly over the unsupervised dictionary, leading to state-of-the-art performance on diverse image databases. Further more, our supervised model targets learning linear features, implying its great potential in handling large scale datasets in real applications.
摘要
针对分类任务,提出了一种新颖的基于局部图像描述子的监督分级稀疏编码模型。
通过 back-projection方法,以最小化在图像层级特征(image level features)的分类误差训练监督词典。其中图像层级特征是以空间金字塔为结构max pooling稀疏编码。在多种空间尺度下max pooling方法具有平移不变的特性,如同 CNN(Convolutional Neural Network)一样。
实验证明,与无监督词典相比,监督词典明显地改善了模型的性能,并且在多个图像数据库拥有最好的表现。
另外,监督模型目标是学习线性特征,它蕴含了一个巨大潜能 -实时地处理大规模数据库。
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
拟解决的问题
Image classification To find a generic feature representation Interested in linear prediction model
Sparse Coding for Image Classification
Sparse Coding Unsupervised Supervised
Sparse coding on holistic image
-Linear model assumption
-Sensitive to image misalignment
-Limited applications
D. Bradley et al. ‘08
J. Wright et al. ’09
A. Wagner et al.’09
etc
D. Bradley et al. ‘08
J. Marialet al. ’08
Q. Zhang. CVPR10
etc
Sparse coding on local descriptors
-Break linear model assumption for the image space
-Robust to image misalignment
-Applicableto generic image
classification
R. Rainaet al. ’07
J. Yang et al. ’09
J. Yang et al. ’10
etc
?
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
本文的方法框架相关知识本文模型求解方法
框架
Bag of coordinatedLocal descriptors
High-dimensionalsparse codes
Imagerepresentation
It must be a cool Cat!
Descriptor extraction
nonlinear coding
feature pooling
classification
J. Yang et al. Linear spatial pyramid matching using sparse coding for image classification. CVPR’09.
Yang. CVPR09
已有方法
Histogram-based SPM feature Step 1: local descriptor extraction Step 2: vector quantization (e.g.k-means) Step 3: hierarchical average pooling Step 4: nonlinear SVM
The framework of ScSPM ( CVPR09) Step 1: local descriptor extraction Step 2: sparse coding (无监督词典 ) Step 3: hierarchical max pooling Step 4: linear SVM
相关知识 (1)
Sparse coding
Max pooling
Xnxm=(X1,X2,…,Xm)
Bnxk:词典
Zkxm:稀疏系数
相关知识 (2)
分级融合
S: 尺度(层次)
U: 串接
Model ( 1)
多层max pooling
+ SVM
目标函数
Xk:表示第 k个图像
监督
Model ( 2 ) -目标函数
Optimization over B: back propagation!
求解方法( 1)
Squared hinge loss function
Linear prediction model
Only cares about the pooled maximum values
No analytical link
求解方法( 2)
Solution: use implicit differentiation
D. M. Bradley et al. Differentiable sparse coding. NIPS 2008.
Setting the gradients at zero coefficients to be zero,
a lot of computations can be saved!
Training convergence
Initialization is important: B is trained in unsupervised manner.
Convergence
Example dictionary
Example dictionary: CMU PIE
Unsupervised Supervised
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
Experiment
Classification tasks Face recognition: CMU PIE, and CMU Multi-PIE Handwritten digit recognition: MNIST Gender Recognition: FRGC 2.0
Image local descriptors: raw image patches Prediction model: one-vs-all linear SVM with squared hinge loss
function. Stochastic optimization: typically converges in 10 iterations,
gradient descent.
Experiment
Parameter settings
学习率:
Experiment –Face Recognition (1)
CMU PIE: 41368 images of 68 people, each under 13 poses, 43
different illumination conditions with 4 different expressions.
A subset of five near frontal views are used including all expressions and illuminations.
Experiment –Face Recognition (1)
USC: unsupervised sparse coding model. SSC: supervised sparse coding model. Improvements: shows the improvements of SSC over
USC.
Classification error(%) on CMU PIE
Experiment –Face Recognition (2)
CMU Multi-PIE: contains 337 subjects across simultaneous variations
in pose, expression and illumination. A subset containing near frontal view face images are
used as training and testing.
Experiment –Face Recognition (2)
[SR] A. Wagner et al. Towards a practical face recognition system: robust registration and illumination by sparse representation. CVPR’09.
Face recognition error(%) on Multi-PIE
Experiment – Handwritten Digit Recognition
MNIST: consists of 70,000 handwritten digits, aligned to the center. 60,000 of them are modeled as training, and the rest 10,000 as testing.
Experiment – Gender Recognition
FRGC 2.0 contains 568 individuals, totally 14714 face images
under various lighting conditions and backgrounds. 11700 face images of 451 individuals are used as
training, and the remaining 3014 images of 114 persons are used as testing.
Experiment – Gender Recognition
提纲
•作者信息•文章信息•拟解决的问题•本文的方法•实验•结论
Conclusion
A supervised translation-invariant sparse coding model for image classification A generic image representation. The max pooling feature is translation-invariant. Sparse coding on local descriptors is promising compared to
sparse coding on holistic image. Supervised sparse coding improves the performance
significantly. Next steps:
Connections with hierarchical models in deep belief networks should be investigated.
More theoretical analysis for pooling functions are needed. Deep hierarchical models based on sparse coding should be
studied.
参考文献
Jianchao Yang, Kai Yu, Thomas Huang,Supervised Translation-Invariant Sparse Coding. CVPR10.
J. Yang et al. Translation-Invariant Sparse Coding. CVPR10(talk). J. Yang et al. Linear spatial pyramid matching using sparse coding
for image classification. CVPR’09.