|
JPEG has been a widely recognized image compression standard for many years. Nevertheless, it faces its own
limitations as compressed image quality degrades significantly at lower bit rates. This limitation has been addressed in
JPEG2000 which also has a tendency to replace JPEG, especially in the storage and retrieval applications. To efficiently
and practically index and retrieve compressed-domain images from a database, several image features could be extracted
directly in compressed domain without having to fully decompress the JPEG2000 images. JPEG2000 utilizes wavelet
transform. Wavelet transforms is one of widely-used to analyze and describe texture patterns of image. Another
advantage of wavelet transform is that one can analyze textures with multiresolution and can classify directional texture
pattern information into each directional subband. Where as, HL subband implies horizontal frequency information, LH
subband implies vertical frequency information and HH subband implies diagonal frequency. Nevertheless, many
wavelet-based image retrieval approaches are not good tool to use directional subband information, obtained by wavelet
transforms, for efficient directional texture pattern classification of retrieved images. This paper proposes a novel image
retrieval technique in JPEG2000 compressed domain using image significant map to compute an image context in order
to construct image index. Experimental results indicate that the proposed method can effectively differentiate and
categorize images with different texture directional information. In addition, an integration of the proposed features with
wavelet autocorrelogram also showed improvement in retrieval performance using ANMRR (Average Normalized
Modified Retrieval Rank) compared to other known methods.
|
