Non-negative Matrix Factorization for Multimodal Image Retrieval

Transcription

1 Non-negative Matrix Factorization for Multimodal Image Retrieval Fabio A. González PhD Bioingenium Research Group Computer Systems and Industrial Engineering Department Universidad Nacional de Colombia Knowledge Discovery and Webmining Lab CECS, The U of L Graduate Seminar, Spring 2010 CECS Department, The University of Louisville F. González (UNAL) NMF for MM IR CECS Grad Sem 1 / 49

2 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 2 / 49

3 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 3 / 49

4 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 4 / 49

5 Content-Based Image Retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 5 / 49

6 Query by Visual Example F. González (UNAL) NMF for MM IR CECS Grad Sem 6 / 49

7 Low-level vs. High-level F. González (UNAL) NMF for MM IR CECS Grad Sem 7 / 49

8 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 8 / 49

9 Semantic Annotation using ML Source: Nuno Vasconcelos, UCSD, F. González (UNAL) NMF for MM IR CECS Grad Sem 9 / 49

10 An Example (1) F. González (UNAL) NMF for MM IR CECS Grad Sem 10 / 49

11 An Example (2) 1 Recall vs Precision Graph - Semantic models Histogram Intersection MetaFeatures Identity Kernel Sobel Precision Recall F. González (UNAL) NMF for MM IR CECS Grad Sem 11 / 49

12 Disadvantages Requires a training set with expert annotations, so it is a costly process F. González (UNAL) NMF for MM IR CECS Grad Sem 12 / 49

13 Disadvantages Requires a training set with expert annotations, so it is a costly process Does not scale for large semantic vocabularies F. González (UNAL) NMF for MM IR CECS Grad Sem 12 / 49

14 Disadvantages Requires a training set with expert annotations, so it is a costly process Does not scale for large semantic vocabularies The mapping from visual features to annotations may lose the visual richness F. González (UNAL) NMF for MM IR CECS Grad Sem 12 / 49

15 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 13 / 49

16 Multimodality Text and images come naturally together in many documents Academic papers, books Newspapers, web pages Medical cases F. González (UNAL) NMF for MM IR CECS Grad Sem 14 / 49

17 Multimodal Retrieval Unstructured text associated to images may be used as semantic annotations Images and texts are complimentary information units Take advantage of interactions between both data modalities F. González (UNAL) NMF for MM IR CECS Grad Sem 15 / 49

18 Multimodal Retrieval Unstructured text associated to images may be used as semantic annotations Images and texts are complimentary information units Take advantage of interactions between both data modalities Problems: Text associated to images is not structured Unclear relationships between keywords and visual patterns Possible presence of noise in both data modalities F. González (UNAL) NMF for MM IR CECS Grad Sem 15 / 49

19 Multimodal Retrieval Unstructured text associated to images may be used as semantic annotations Images and texts are complimentary information units Take advantage of interactions between both data modalities Problems: Text associated to images is not structured Unclear relationships between keywords and visual patterns Possible presence of noise in both data modalities Retrieval scenarios: Cross-modal: find images based on a text query find text based on an image query (image annotation) Visual retrieval based on a visual query F. González (UNAL) NMF for MM IR CECS Grad Sem 15 / 49

20 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 16 / 49

21 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 17 / 49

22 The Netflix Competition Problem: prediction of user ratings for films (collaborative filtering) Y. Koren, R. Bell, and C. Volinsky, Matrix factorization techniques for recommender systems, Computer, vol. 42, no. 8, pp , August 2009 F. González (UNAL) NMF for MM IR CECS Grad Sem 18 / 49

23 The Netflix Competition Problem: prediction of user ratings for films (collaborative filtering) Prize: $1 Million to the first algorithm able to improve Netflix own algorithm results by at least 10% Y. Koren, R. Bell, and C. Volinsky, Matrix factorization techniques for recommender systems, Computer, vol. 42, no. 8, pp , August 2009 F. González (UNAL) NMF for MM IR CECS Grad Sem 18 / 49

24 The Netflix Competition Problem: prediction of user ratings for films (collaborative filtering) Prize: $1 Million to the first algorithm able to improve Netflix own algorithm results by at least 10% Won on Sept/21/2009 by BellKor s Pragmatic Chaos Y. Koren, R. Bell, and C. Volinsky, Matrix factorization techniques for recommender systems, Computer, vol. 42, no. 8, pp , August 2009 F. González (UNAL) NMF for MM IR CECS Grad Sem 18 / 49

25 The Netflix Competition Problem: prediction of user ratings for films (collaborative filtering) Prize: $1 Million to the first algorithm able to improve Netflix own algorithm results by at least 10% Won on Sept/21/2009 by BellKor s Pragmatic Chaos Their approach used Non-negative Matrix Factorization (NMF) to build a latent-factor representation of users and movies Y. Koren, R. Bell, and C. Volinsky, Matrix factorization techniques for recommender systems, Computer, vol. 42, no. 8, pp , August 2009 F. González (UNAL) NMF for MM IR CECS Grad Sem 18 / 49

26 The Latent-Factor Model movies factors movies r r 1m q q 1f p p 1m users users..... = r n1... r nm q n1... q nf p f 1... p fm factors R QP Q, P 0 n m {(i, j) r ij 0} 10 8 f 200 F. González (UNAL) NMF for MM IR CECS Grad Sem 19 / 49

27 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 20 / 49

28 Non-negative Matrix Factorization Problem: to find a factorization Optimization problem: X n m = W n r H r m min A,B X WH 2 s.t. W, H 0 is the Frobenius norm It is a non-convex optimization problem Solution alternatives: Gradient descendent methods Multiplicative updating rules F. González (UNAL) NMF for MM IR CECS Grad Sem 21 / 49

29 Multiplicative Rules Optimization problem: min A,B X WH 2 s.t. W, H 0 Incremental optimization: H aµ H aµ (W T X ) aµ (W T WH) aµ W ia W ia (XH T ) αµ (WHH T ) αµ F. González (UNAL) NMF for MM IR CECS Grad Sem 22 / 49

30 Divergence Optimization Optimization problem: min A,B D(X WH) = ij s.t. W, H 0 Multiplicative Rules: ( X ij log ) X ij (WH) ij X ij + (WH) ij i H aµ H W iax iµ /(WH) iµ aµ i W ia µ W ia W H aµx iµ /(WH) iµ ia µ H aµ F. González (UNAL) NMF for MM IR CECS Grad Sem 23 / 49

31 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 24 / 49

32 PCA and SVD Problem: X n m = W n r H r m Principal Component Analysis (PCA) X = UΣV W = UΣ 1 2, H = Σ 1 2 V PCA = SVD keeping the best Eigenvectors Columns of U are orthonormal There is not restriction on sign. D. D. Lee and H. S. Seung, Learning the parts of objects by non-negative matrix factorization, Nature, vol. 401, no. 6755, pp , October 1999 F. González (UNAL) NMF for MM IR CECS Grad Sem 25 / 49

33 Latent Semantic Indexing (LSI) documents factors documents x x 1m w w 1r h h 1m terms terms..... = x n1... x nm w n1... w nr h r1... h rm X W H factors Documents are represented by the frequency of keywords (terms) Uses SVD to find the factorization Factors = semantic concepts Columns of W are orthonormal F. González (UNAL) NMF for MM IR CECS Grad Sem 26 / 49

34 NMF vs LSI W. Xu, X. Liu, and Y. Gong, Document clustering based on non-negative matrix factorization, in SIGIR 03: Proceedings of the 26th annual international ACM SIGIR conference on Research and development in information retrieval. New York, NY, USA: ACM, 2003, pp F. González (UNAL) NMF for MM IR CECS Grad Sem 27 / 49

35 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 28 / 49

36 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 29 / 49

37 Semantic Space: Multimodal Latent Indexing (I) Objects are described by terms in a textual vocabulary and a visual vocabulary F. González (UNAL) NMF for MM IR CECS Grad Sem 30 / 49

38 Semantic Space: Multimodal Latent Indexing (I) Objects are described by terms in a textual vocabulary and a visual vocabulary Objects are mapped to a latent (semantic) space where objects are represented by a set of latent factors F. González (UNAL) NMF for MM IR CECS Grad Sem 30 / 49

39 Semantic Space: Multimodal Latent Indexing (I) Objects are described by terms in a textual vocabulary and a visual vocabulary Objects are mapped to a latent (semantic) space where objects are represented by a set of latent factors NMF is used to build the latent representation F. González (UNAL) NMF for MM IR CECS Grad Sem 30 / 49

40 Semantic Space: Multimodal Latent Indexing (I) Objects are described by terms in a textual vocabulary and a visual vocabulary Objects are mapped to a latent (semantic) space where objects are represented by a set of latent factors NMF is used to build the latent representation Three main tasks: Multimodal clustering Automatic image annotation Image retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 30 / 49

41 Semantic Space: Multimodal Latent Indexing (II) F. González (UNAL) NMF for MM IR CECS Grad Sem 31 / 49

42 Bag-of-Features Image Representation F. González (UNAL) NMF for MM IR CECS Grad Sem 32 / 49

43 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 33 / 49

44 Multimodal Clustering F. González (UNAL) NMF for MM IR CECS Grad Sem 34 / 49

45 Dual Multimodal Clustering F. González (UNAL) NMF for MM IR CECS Grad Sem 35 / 49

46 Does It Work? Clustering Corel data set: 1000 images, 25 categories Input matrix: 2500 vectors of dimension 1000 Clustering performance comparison: K-means and two NMF variants: K-means NMF (Frobenius norm) NMF (KL divergence) Accuracy *results are average of 10 runs F. González (UNAL) NMF for MM IR CECS Grad Sem 36 / 49

47 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 37 / 49

48 Image Annotation F. González (UNAL) NMF for MM IR CECS Grad Sem 38 / 49

49 Does it Work? plsa has shown good performance for image annotation NMF with KL divergence was shown to be equivalent to plsa W. Xu, X. Liu, and Y. Gong, Document clustering based on non-negative matrix factorization, in SIGIR 03: Proceedings of the 26th annual international ACM SIGIR conference on Research and development in information retrieval. New York, NY, USA: ACM, 2003, pp F. González (UNAL) NMF for MM IR CECS Grad Sem 39 / 49

50 Outline 1 The Problem Content-based image retrieval Semantic image retrieval Multimodal image retrieval 2 Non-Negative Matrix Factorization The Netflix Prize Non-negative matrix factorization NMF vs SVD 3 NMF for Multimodal Retrieval Semantic space Multimodal clustering Image annotation Multimodal retrieval F. González (UNAL) NMF for MM IR CECS Grad Sem 40 / 49

51 Image Retrieval Scenario: visual retrieval based on a visual query F. González (UNAL) NMF for MM IR CECS Grad Sem 41 / 49

52 Image Retrieval Scenario: visual retrieval based on a visual query Images are represented in a latent space using NMF F. González (UNAL) NMF for MM IR CECS Grad Sem 41 / 49

53 Image Retrieval Scenario: visual retrieval based on a visual query Images are represented in a latent space using NMF Some images in the database may not have text content associated F. González (UNAL) NMF for MM IR CECS Grad Sem 41 / 49

54 Image Retrieval Scenario: visual retrieval based on a visual query Images are represented in a latent space using NMF Some images in the database may not have text content associated The image query is projected the latent space F. González (UNAL) NMF for MM IR CECS Grad Sem 41 / 49

55 Image Retrieval Scenario: visual retrieval based on a visual query Images are represented in a latent space using NMF Some images in the database may not have text content associated The image query is projected the latent space Image are retrieved according to their latent space similarity F. González (UNAL) NMF for MM IR CECS Grad Sem 41 / 49

56 Experimental Evaluation Corel Images A subset of 2,500 images extracted from the Corel Database. 25 categories with 100 images each. F. González (UNAL) NMF for MM IR CECS Grad Sem 42 / 49

57 Bag of Features Corel Images Image content representation using parts of images Blocks are extracted from each image and the SIFT descriptor is computed A dictionary of visual patterns is built using k-means A histogram counting the occurrence of each codeblock is constructed for each image F. González (UNAL) NMF for MM IR CECS Grad Sem 43 / 49

58 Results This approach outperforms direct image matching by almost 4X F. González (UNAL) NMF for MM IR CECS Grad Sem 44 / 49

59 Results Some specific performance measures Measure NMF Search P1 Direct Search P P P R R R MAP F. González (UNAL) NMF for MM IR CECS Grad Sem 45 / 49

60 Semantic Space Visualization (I) F. González (UNAL) NMF for MM IR CECS Grad Sem 46 / 49

61 Semantic Space Visualization (II) F. González (UNAL) NMF for MM IR CECS Grad Sem 47 / 49

62 Ongoing Work Bigger data sets: ImageClefMed, Corel5000, Flickr Different image low-level features Kernel NMF Incremental NMF F. González (UNAL) NMF for MM IR CECS Grad Sem 48 / 49

63 Thanks! F. González (UNAL) NMF for MM IR CECS Grad Sem 49 / 49