Image Slicing and Statistical Layer Approaches for Content-Based Image Retrieval
Two new approaches for colour features representation and comparison in digital images to handle various problems in the field of content-based image retrieval are proposed. The first approach is a double-layered system utilising a new technique, which is based on image slicing, combined with stati...
Saved in:
Main Author: | |
---|---|
Format: | Thesis |
Language: | English English |
Published: |
2004
|
Subjects: | |
Online Access: | http://psasir.upm.edu.my/id/eprint/8/1/100548937_t_FSKTM_2004_10.pdf |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Summary: | Two new approaches for colour features representation and comparison in digital images to handle various problems in the field of content-based image retrieval are proposed.
The first approach is a double-layered system utilising a new technique, which is based on image slicing, combined with statistical features extracted and compared in each layer (ISSL). The images database is filtered in the first layer based on the similarities of brightness compared with the query image and ranked in the second layer, based on the
similarities of the contrast values between the query image and the set of candidate images retrieved through the first layer. Although different distance measurements are
available, the city block known as L1-norm distance measurement is used. This is due to its speed efficiency and accuracy. Different experiments are applied to different database sets, containing different number of images. The results show that the approach is scalable to the varying size of the database, robust, accurate, and fast. A comparison between the colour histogram approach and the proposed approach shows that the proposed system is more accurate and the speed of performance is much better. A new
paradigm to choose the proper threshold value is proposed based on the autocorrelation of the distance vector. Moreover, an image retrieval system based on entropy as a visual discriminator is developed and compared with ISSL. The results show that the proposed ISSL approach is able to achieve better precision and reaches higher recall levels as
compared with entropy approach.
The second proposed technique for colour based retrieval is the Eigenvalues approach. Findings show that the interpretation of the Eigenvalues, as identity or signature for the square matrix, makes it possible to map this concept to the different bands of the image. The approach relies on calculating the accumulative distances between the query image and the images database, using the accumulative Eigenvalues of each band. The approach is tested, using different image queries over different database sets and the results are promising. Furthermore, the proposed approach is compared with ISSL approach and entropy approach, using different query images over a database set of 2000 images.
In addition, a shape-based retrieval system is proposed. The system is double-layered, in which the first layer is used to filter the images database based on colour similarity. This
allows the reduction in the number of candidate images, which need to be manipulated, using the shape retrieval technique in the second layer. The technique utilises a low-level image processing operations with “Dilate” as a morphological operator. Laplacian of Gaussian (LoG) is used to smoothen and detect the edges of the objects. Dilate on the other hand is used to solidify the object and fill in the holes, and correlation coefficient is proposed as a new means to shape similarity measurement. Experiments show that the approach is fast, flexible, and the retrieval of images is highly accurate. It is also able to overcome the numerous problems that are associated with the usage of the low-level image processing operation in image retrieval.
|
---|