Relevance Feedback-Based Image Retrieval Interface Incorporating Region and Feature Saliency Patterns as Visualizable Image Similarity Criteria Speaker.

Relevance Feedback-Based Image Retrieval InterfaceIncorporating Region and Feature Saliency Patterns

as Visualizable Image Similarity Criteria

Speaker： Kun Hsiang

Outline

RFSP Image Similarity Model

• GA-Based Relevance Feedback Mechanism Using RFSP

• Experimental Evaluation of RFSP Method


• Set of images I represents the image database.

• The area of each image is partitioned into nR regions, defined by the set of regions R .

• From each region, nF features are extracted, based on the set of features F.

• The set of weights W contains nW real-valued region and feature weights.



• Region is defined as a rectangular array of image pixels, and all image regions are of equal size, obtained by uniformly partitioning the image area into NxN (=nR) blocks.

• Feature denotes an arbitrary image feature , e.g., color or texture ,based on which the similarity of a pair of image regions is computed.

• Weight denotes the relative importance of a region or a feature.






• Thanks to the relevance feedback, all the user has to do is to specify the query image and a couple of relevant images, without worrying about the region and feature weights, or at all being aware of the existence of the RFSP structure.



GA-Based Relevance Feedback Mechanism Using RFSP

• GA for inferring RFSP：– We employed five weight levels：minimal (0),

low(0.25), medium(0.5), high(0.75), maximal (1)

– W={w1,w2,w3,w4,w5}


• GA for inferring RFSP：– An important difference between the proposed

RFSP model and all of the surveyed relevance feedback methods is that the proposed model uses a discrete set of (region and feature) weights, rather than arbitrary weights from interval [0, 1].


• GA for inferring RFSP：– Using a discrete set of weights：

• While theoretically decreasing the expressive power of the model

• Also contributes to the faster convergence of the GA, and does not have practical implications on the retrieval performance, as confirmed through the preliminary experiments.

• Increasing the number of weight levels only made the convergence slower, without improving the retrieval performance.

• In the preliminary experiments, five weight levels resulted in the optimal balance between the speed of GA convergence and the retrieval performance.


• Chromosome coding：– Gene


• Chromosome coding：– Chromosome


• Fitness measure



• Evolution parameters：– Regarding the parameters of the GA evolution,

selection is a standard proportional selection incorporating elitist model.

– Crossover probability is 0.6– Mutation probability is 0.1– Population size is 50 chromosomes– Evolution length is 250 generations.

Experimental Evaluation of RFSP Method

• Test Databases：– Vistex-60 database– Vistex-167 database– Brodatz-208 database– Corel-1000-A database– Corel-1000-B database






GA


LSP


RFSP

Relevance Feedback-Based Image Retrieval Interface Incorporating Region and Feature Saliency Patterns as Visualizable Image Similarity Criteria Speaker.

Documents