23-01-2013, 10:20 AM
Content-based Image Indexing and Searching Using Daubechies' Wavelets
1Content-based Image.pdf (Size: 1.7 MB / Downloads: 80)
Abstract.
This paper describes WBIIS (Wavelet-Based
Image Indexing and Searching), a new image indexing
and retrieval algorithm with partial sketch image search-
ing capability for large image databases. The algorithm
characterizes the color variations over the spatial extent
of the image in a manner that provides semantically-
meaningful image comparisons. The indexing algorithm
applies a Daubechies' wavelet transform for each of the
three opponent color components. The wavelet coe-
cients in the lowest few frequency bands, and their vari-
ances, are stored as feature vectors. To speed up re-
trieval, a two-step procedure is used that rst does a
crude selection based on the variances, and then renes
the search by performing a feature vector match between
the selected images and the query. For better accuracy
in searching, two-level multiresolution matching may
also be used. Masks are used for partial-sketch queries.
This technique performs much better in capturing co-
herence of image, object granularity, local color/texture,
and bias avoidance than traditional color layout algo-
rithms. WBIIS is much faster and more accurate than
traditional algorithms.
Introduction
Searching a digital library [21] having large number of
digital images or video sequences has become important
in this visual age. Every day, large numbers of people are
using the Internet for searching and browsing through
dierent multimedia databases. To make such searching
practical, eective image coding and searching based on
image semantics is becoming increasingly important.
Preprocessing the Images in the Database
Many color image formats are currently in use, e.g., GIF,
JPEG, PPM and TIFF are the most widely used for-
mats. Because images in an image database can have
dierent formats and dierent sizes, we must rst nor-
malize the data. For our test database of relatively small
images, a rescaled thumbnail consisting of 128128 pix-
els in Red-Green-Blue (i.e., RGB) color space is adequate
for the purpose of computing the feature vectors.
Bilinear interpolation is used for the rescaling pro-
cess. This method resamples the input image by overlay-
ing the input image a grid with 128 128 points. This
gives one grid point for each pixel in the output image.
The input image is then sampled at each grid point to de-
termine the pixel colors of the output image. When grid
points lie between input pixel centers, the color values
of the grid point are determined by linearly interpolat-
ing between adjacent pixel colors (both vertically and
horizontally).