Local Binary Pattern -menetelmä kontekstipohjaisessa kuvahaussa (Local Binary Pattern method in context-based image retrieval). Takala, V. Ph.D. Thesis, 2004.
abstract   bibtex   
Textures play an important role in modern context-based image retrieval (CBIR). Using them as image feature descriptors is at least as justified as is the use of color because most pictures contain textures. Many interesting objects would be left unfound without the help of texture information. In this thesis, the suitability of the Local Binary Pattern (LBP) texture methods was considered in context-based image retrieval. During the research two different block-based approaches were developed. The first one of them divides both the query and database images into blocks and compares the image blocks according to their feature histograms. The final distance returned by the algorithm is the total sum of block distances or only a part of it. The second approach uses the block division method to the database images only. It connects the image areas called as basic blocks according to the size of query image and performs the comparison between the query feature vector and the sum of basic block features. The best matching basic block area is found by moving the search window over the whole image area. Color correlogram, which is based on color and works in a kind of similar way to LBP, was used as a reference method. Its performance in context-based image retrieval is state of the art. The performance of the developed block algorithms was evaluated by using a C++ implementation that was build along the work project. There were two test databases: the bigger of them was a heterogeneous image collection composed of Corel pictures and the smaller one was a German stamp database which could also be used through a commercial retrieval software. The LBP block methods succeeded in both databases. The average retrieval results in the Corel database were better than the ones provided by the full image color correlogram, and also the results of the special tests on the stamp database were competitive. The greatest challenges of the researched block algorithms are related to the resource requirements as the numerous feature histograms calculated for the image blocks require a lot of memory and processing capacity. In addition, the changes in texture scale cause problems for the LBP features because they are not invariant to scaling.
@phdthesis{
 title = {Local Binary Pattern -menetelmä kontekstipohjaisessa kuvahaussa (Local Binary Pattern method in context-based image retrieval)},
 type = {phdthesis},
 year = {2004},
 id = {9eaedae3-c94d-30ba-9471-9455087e800f},
 created = {2019-11-19T16:28:42.795Z},
 file_attached = {false},
 profile_id = {bddcf02d-403b-3b06-9def-6d15cc293e20},
 group_id = {28b2996c-b80f-3c26-be71-695caf7040ac},
 last_modified = {2019-11-19T16:32:31.951Z},
 read = {false},
 starred = {false},
 authored = {false},
 confirmed = {true},
 hidden = {false},
 citation_key = {mvg:581},
 source_type = {mastersthesis},
 notes = {M.Sc. thesis, Department of Electrical and Information Engineering, University of Oulu, Finland, 63 p + App.},
 folder_uuids = {8292f5ec-1c57-4113-a303-25778e695f8c},
 private_publication = {false},
 abstract = {Textures play an important role in modern context-based image retrieval
(CBIR). Using them as image feature descriptors is at least as justified as is the
use of color because most pictures contain textures. Many interesting objects
would be left unfound without the help of texture information.

In this thesis, the suitability of the Local Binary Pattern (LBP) texture methods
was considered in context-based image retrieval. During the research two
different block-based approaches were developed. The first one of them divides
both the query and database images into blocks and compares the image blocks
according to their feature histograms. The final distance returned by the algorithm
is the total sum of block distances or only a part of it. The second approach
uses the block division method to the database images only. It connects
the image areas called as basic blocks according to the size of query image and
performs the comparison between the query feature vector and the sum of basic
block features. The best matching basic block area is found by moving the
search window over the whole image area. Color correlogram, which is based
on color and works in a kind of similar way to LBP, was used as a reference
method. Its performance in context-based image retrieval is state of the art.

The performance of the developed block algorithms was evaluated by using a
C++ implementation that was build along the work project. There were two test
databases: the bigger of them was a heterogeneous image collection composed of
Corel pictures and the smaller one was a German stamp database which could
also be used through a commercial retrieval software. The LBP block methods
succeeded in both databases. The average retrieval results in the Corel database
were better than the ones provided by the full image color correlogram, and
also the results of the special tests on the stamp database were competitive.

The greatest challenges of the researched block algorithms are related to the
resource requirements as the numerous feature histograms calculated for the
image blocks require a lot of memory and processing capacity. In addition, the
changes in texture scale cause problems for the LBP features because they are
not invariant to scaling.},
 bibtype = {phdthesis},
 author = {Takala, V}
}

Downloads: 0