1 / 21

A Hybrid Region Weighting Approach for Relevance Feedback in Region-Based Image Search on the Web

A Hybrid Region Weighting Approach for Relevance Feedback in Region-Based Image Search on the Web. Deok-Hwan Kim, Jae-Won Song, Ju-Hong Lee INHA University. Contents. Problem of CBIR Proposed RBIR Hybrid Region Weighting Experiment and Results Conclusion. Content Based Image Retrieval.

arion
Télécharger la présentation

A Hybrid Region Weighting Approach for Relevance Feedback in Region-Based Image Search on the Web

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. A Hybrid Region Weighting Approach for Relevance Feedback in Region-Based Image Search on the Web Deok-Hwan Kim, Jae-Won Song, Ju-Hong Lee INHA University

  2. Contents • Problem of CBIR • Proposed RBIR • Hybrid Region Weighting • Experiment and Results • Conclusion

  3. Content Based Image Retrieval • CBIR • utilizes unique features (shape, color, texture) of images Users prefer • To retrieve relevant image by semantic categories • But, CBIR can not capture high-level semantics in user’s mind

  4. Problem of CBIR (1) • Problem • Focused on developing effective global features  Can not capture properties of an object The gap between low-level feature and high-level semantics System Semantic Gap Query Image User

  5. Problem of CBIR(2) • Solutions • Relevance Feedback (RF) • Region-based Image Retrieval (RBIR)

  6. Relevance Feedback • Relevance Feedback • Learns the associations between high-level semantics and low-level features • Relevance Feedback Phase • User identifies relevant images within the returned set • System utilizes user feedback in the next round • To modify the query (to retrieve better results) • This process repeats  until user is satisfied

  7. Problem of CBIR(2) • Solutions • Relevance Feedback (RF) • Region-based Image Retrieval (RBIR)

  8. RBIR(Region-Based Image Retrieval) • Region-Based approaches • Represent image at the object level • The main objective • Enhance the ability of capturing user’s perception • More meaningful retrieval • Image similarity measures • EMD • Weighting of region • key factor of similarity definition.

  9. Image Top k Retrieved Set Segmentation User Feedback Loop Image DB Region 1 Region n Relevant Set Feature extraction Feature extraction Adaptive Clustering Weight Computation Weight Computation Q=(q,d,w, k) ClusterRepresentatives Region basedCluster Set Q =(q,d,w, k) EMD match Proposed RBIR Approach

  10. Adaptive Region Clustering(1) • Merges similar regions in the relevant set  reduce retrieval speed • T2 > Threshold : separate two clusters • T2 <=Threshold : merge two clusters

  11. (g-1)th level C1 Cg-1 C1 Ck Cg gth level C1 C2 Cm Cn Region Image Adaptive Region Clustering(2)

  12. Image Segmentation and Region Representation • Normalized cut segmentation • Discriminate foreground object regions and background regions • Region Representation in an Image • Twelve dimensional color and shape features • Color feature • mean, standard deviation of color in L*a*b color space • Shape feature • Compactness and convexity, region size, region location, and variance of region pixels from the region center of mass

  13. Region Weighting • Existing Region Weighting • Area Percentage • RF (Region Frequency) * IIF (Inverse Image Frequency) • Suggested Region Weighting • spatial locations of regions • region size in an image

  14. S xn x1 RB x3 RA x2 Center C random points l RC Hybrid Region Weighting • Assume that more important region • appear in center area of an image • tend to occupy larger area • To consider image’s Spatial location

  15. Hybrid Region Weighting • Region Importance • Calculated by summarizing the reciprocal function values with respect to all pixel locations x of region Rki • However, it is difficult • Instead, use the asymptotic distance function by applying the Monte-Carlo method

  16. Hybrid Region Weighting • Region Weight • Decay Factor β (0≤ β ≤1) • reduce the effect of previous relevant image • We assume that there are n relevant images I1…In • prior images : I1…Im • new images : Im+1… In

  17. Hybrid Region Weighting • New region weights using decay factor is as follows:

  18. Three weights for regions of an animal image area percentage 0.16 region frequency 0.10 area & location 0.06 area percentage 0.01 region frequency 0.11 area & location 0.07 area percentage 0.06 region frequency 0.12 area & location 0.12 area percentage 0.24 region frequency 0.04 area & location 0.42 area percentage 0.06 region frequency 0.23 area & location 0.18 area percentage 0.47 region frequency 0.39 area & location 0.15

  19. Experiment and Results(1) • k-NN query • used to accomplish the similarity-based match • k = 100. • For RBIR with RF approach • Use adaptive region clustering method • 10,000 general purpose color images from COREL • 40 random initial query • five feedback • For decay factor, empirically, β =0.3 • To evaluate the performance, we compare • Area percentage, Region frequency, Area & location

  20. Experiment and Results(2) Performance evaluation

  21. Conclusion • The main contribution • Calculate the importance of regions by using the hybrid weighting method • Cumulate it based on user’s feedback information to better represent semantic importance of a region in a given query • Proposed weighting method can also be incorporated into any RBIR system on Web • It put more emphasis on the latest relevant images that express the user’s query concept more precisely • Experimental results • Show the superiority of the proposed method over other weighting methods in terms of efficiency and effectiveness

More Related