1 / 30

Enhancement of textual images classification using their global and local visual content

Enhancement of textual images classification using their global and local visual content. Sabrina Tollari, Hervé Glotin, Jacques Le Maitre Université de Toulon et du Var Laboratoire SIS - Équipe Informatique France MAWIS 2003. Plan. Objective State of the Art Presentation of the corpus

matty
Download Presentation

Enhancement of textual images classification using their global and local visual content

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Enhancement of textual images classificationusing their global and localvisual content Sabrina Tollari, Hervé Glotin, Jacques Le Maitre Université de Toulon et du Var Laboratoire SIS - Équipe Informatique France MAWIS 2003

  2. Plan • Objective • State of the Art • Presentation of the corpus • Presentation ot the system • 3 experiments • Conclusion

  3. Visual filter Enhancement of image search engine

  4. Textual indices Visual features Paysage Cameroun Agriculture Complementarity The problem

  5. Textual indices Low or High level visual features • Textual indices : • Manual indexation, Keywords… • Auto indexation : Legend, surrounding text… • Low level Visual features : • - Color : RGB, HSV, Brightness, Edges • High level Visual features : • - Shape, Spectrum analysis (rotation invariance) Fourrier transform, wavelet,texture - Semantics of textual indice >> semantics of High Level Features - Low level visual features requieres low computation time and give complementarity information to textual indices

  6. State of the art: text OR visual information ? * None of them is using text information to enhance visual querying and vice versa ** Early fusion merging textual and visual features, or visual / textual user feedback

  7. The corpus (1/2) • 600 press agency photos • Textually indexed by a picture researcher with keywords from a hierarchical thesaurus • Stored as MPEG-7 descriptions in an XML file

  8. The corpus (2/2) • Simple automatic low-level visual features (histograms) • Red • Green • Blue • Brightness • Edge direction

  9. 50% 50% Step B Ramdom division Reference database Test database with a priori semanti classes Method Corpus Step A Semantic textual classification Step C Evaluation of an automatic classification of the test database using the reference database

  10. Step A : Construction of a textual semantics using an ascendant hierarchical classification • Lance & Williams, 1967 • Objective : cluster similar images • Highlight of non trivial semantic classes • Check the class cardinal

  11. Média(1) Radio(2) Télévision(3) Thésaurus Vectorial representation • Salton, 1971 • Ex : Let I be the image such that Term(I)={Radio} • Vector(I)=(0,1,0) • Extended_vector(I)=(1,1,0)

  12. New agregation criterion One element of class A CT dist(X,Y) = 1- | Class B Similarity measure for the AHC Let x et y the vectors of the images X et Y | • Classical criterion • nearest neighbour • farthest neighbour

  13. Step A : construction of the semantic classes • 24 classes • Each contains 8 to 98 images 3 most frequent keywords of some classes :

  14. 50% 50% Step B Ramdom division Reference database Test database with a priori semanti classes Method Corpus Step A Semantic textual classification

  15. Evaluated class Ce C1 Paysage, agriculture, Cameroun C2 Image of the test database (original class Co) Femme, Ouvrier, Industrie Reference database Step C: Evaluation of an automatic classification of the test database using the reference database If Co¹Cethen classification error

  16. Kullback-Leibler distance (1951) Let x and y be two probability distributions Kullback-Leibler divergence: Kullback-Leibler distance: Step C: 3 different classifications • Text Only Classification • Visual Only Classification • Text and Visual Classification

  17. 1.Text Only Classification 1. Results of Text Only Classification • Average vector for each class • Textual class of the image IT:

  18. IT Test image N=2 I1 0.2 Reference class Ck I2 0.6 d(IT,Ck)=0.25 I3 0.3 Average of the N first minimal distances I4 0.8 2.Visual Only Classification Early fusion of visual features

  19. 2.Visual Only Classification Results of Visual Classification * Error rate in % Theoritical error rate: 91.6%

  20. 3.The late visuo-textual fusion The late visuo-textual fusion • Evaluation of the probability that image IT belongs to class Ck by late visuo-textual fusion

  21. 3.The late visuo-textual fusion Class Probability definitions A Î{Red, Green, Blue, Brightness, Edge direction}

  22. 3.The late visuo-textual fusion Weighting definition • Let TE(j) be the error rate using visual features Aj • Weighting distorsion using power p

  23. 3.The late visuo-textual fusion Result : Enhancement of textual classificationincreases with p Textual only Enhancement Visuo-textual Visual Only Error Rate : 71 %

  24. 3.The late visuo-textual fusion Summary of the visuo-textual enhancementusing global content Low-level visual features improve textual classification.

  25. Conclusion • We presented a simple system for unifying textual and visual informations. • We showed that visual information reduces the errors of the textual information without thesaurus of about 50% • Our corpus being only of 600 images, our method must be tested on a database of more significant data.

  26. Discussion • Other visual attributes as texture or form could be used. • Many criteria and parameters remain to be studied to improve visual description, as the influence of the size of the visual content . • Our system can be added as fast visual filter on the result of a request of images on a search engine (such as Google).

  27. Automatic indexing : •  economy • dollars • family • people ? Perspectives • One can reverse the experiment to correct a bad textual indexing using the visual content.

  28. Thank you for your attention

  29. Discussion Local visual content and Region of interest

  30. Enhancement results of late fusions forlocal and global visual content

More Related