Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 83

In practice, the background region or other non-relevant regions may be partially
included in S . These non-relevant regions have a large area outside the W m × n

window. In these cases, a region S if its area outside

the window is greater than its area inside the window. Thus, a complete region of
interest is defined as:

R i , i ∈ W , is removed from

≡ S

c
i ∗

ROI

i ∗ R

(3.29)

c

where R i ∗ is the region that has the sum of pixels

located outside the W m × n window greater than the sum of pixels located inside the
window.
Figure 3.10 shows some image examples after applying the ROI characterization
process to four images from the Corel digital collection.

R

i ∗ is the complement of

R i ∗ , and

3.4.4 Pseudo-Relevance Feedback with Region of Interest

Figure 3.11 shows a diagram of the automatic adaptive retrieval system with
embedding knowledge of ROI. Automatic retrieval is performed in two stages.
First, the SOTM is applied for relevance classification to label the retrieved samples
as positive or negative. Then, the labeled samples form the input to a non-linear
similarity ranking function based on a single-class RBF network. Here, the query
and the retrieved samples are each represented by a point in the feature space F 1 and

consist of color, shape, and texture features. As discussed previously, in order for the
unsupervised learning process brought into automatic RF to be effective, a different
and more powerful feature space other than

F 1 should be introduced in relevance

classification. Thus, the feature space F 2 is used. Apparently features extracted from

the ROIs satisfy this requirement as they provide the embedded knowledge of ROI to
assist in relevance classification. Color and shape are again chosen as the features,
but they are calculated only from the ROIs in the retrieved images after applying
ROI identification.

3.4.5 Experimental Result

Results provided in this section were obtained using the Corel Digital Library, which
contains more than 11,500 photos. Each image is indexed by 48-bin HSV color
histograms and color moments, using the Gabor wavelet method and the Fourier
descriptor [ 97 ]. This produces a 115-dimensional feature vector in the feature space
F 1 which is used for retrieval. For the identification of image relevancy by SOTM,
the retrieved images are passed through the Edge Flow algorithm to identify ROIs,