Self-adaptation in Image and Video Retrieval - Multimedia Database Retrieval: Technology and Applications - page 68

Table 3.1 Retrieval results, obtained by the automatic retrieval system
using pseudo-RF and self-organization methods
Classifier Query Set A Query Set B Query Set C

Initial retrieval 37.8 39.2 39.8

SOM 51.2 49.1 52.3

SOTM 52.1 50.6 54.4

SOM with GA weighting 61.2 59.9 63.6

SOTM with GA weighting 66.8 65.7 63.3

operation), for all three sets of queries. It can be observed that the application of
pseudo-RF employing the two self-organizing methods improves retrieval accuracy
significantly. This result also shows that the map obtained by SOTM is more suitable
for pseudo-labeling, as compared to that of SOM.
In addition, the feature maps obtained by SOTM and SOM can be further
improved by incorporating feature weighting algorithm [ 75 ]. In this case, the
Genetic Algorithm (GA) was applied to the feature set in F 2 to automatically assign

different weights to the color, shape, and texture features, during the construction of
the maps. The experimental results are illustrated in the last two rows of Table 3.1 .
The self-organizing methods offer even higher accuracy with the application of the
GA feature weighting algorithm.

3.3 Re-ranking in Compressed Domains

Compressed domain descriptors are widely used in image retrieval [ 76 - 79 ]. The
main reason for this is that feature extraction algorithms can be applied to image
and video databases without full decompression. This provides fast and efficient
tools that are appropriate for real-time applications. In this section, the compressed
domain features are considered as feature space

F 1 for database retrieval. The
characteristics of discrete cosine transform (DCT) in the Joint Photographic Experts
Group (JPEG) standard and discrete wavelet transform (DWT) in image compres-
sions are investigated for adaptive retrieval in compressed domains.

3.3.1 Descriptor in Discrete Cosine Transformation

Largely due to its energy packing property, DCT has been widely used in JPEG and
many other popular image and video compression standards. When a typical 8 8

block of data undergoes DCT transformation, most of the significant coefficients are
concentrated in the upper-left (low frequency) region of the transform block, thus
allowing the storage and transmission of a small number of coefficients.

×