We present a system for example-based image database retrieval with automatically weighted, selective features and show results of the system on a database of 3000 images.
Many image database systems use un-selective global features such as color histograms. Queries are limited to a single positive example image and the user is required to manually weight the relative importance of features.
Our approach measures the statistical properties of an image such as response energies to local edge filters at various orientations. These responses are in turn used as input to another level of filtering. The resulting filtering sequence captures the response energy of the image to a particular selective feature. Our system allows the user to select both multiple positive and negative example images.
Boosting is used to combine a set of weak classifiers into a strong ranking engine.
Feature Extraction Engine
Strong
Ranking
Engine
feature vectors
Overview
Weak Classifier
Weak Classifier
Weak Classifier
Weak Classifier
ranked images
user-selected
examples
image database
Boosting
Results
Feature Selection
Although a large number of features provides a rich representation, it also requires a large amount of computation time and memory. In addition, some features may not lead to useful projections. We’ve selected features with high kurtoses across a sample of images. This chooses features which are highly selective and hence lead to more interesting non-Gaussian projections.
Histograms of kurtotic features
Histograms of random features
Textures-of-Textures
Feature Extraction
feature vector
Boosting
weak classifier 1
initial equal weight distribution of training examples
weak classifier 2
incorrect classifications re-weighted more heavily
further re-weighting
weak classifier 3
final classifier is weighted combination of weak classifiers
BOOSTING IMAGE RETRIEVAL
Kinh Tieu and Paul Viola
Back to first slide | View graphic version |