Image retrieval using compact Binary signatures

By | May 10, 2013


Significant research has focused on determining efficient methodologies for retrieving images from large image databases. The thesis addresses the concept of new image abstraction technique based on compact binary signatures. The enormous growth of image archives has significantly increased the demand for research efforts aimed at efficiently finding similar images within large image databases. Though there are a number of image retrieval techniques available like image retrieval using M Band Wavelet Transform, Object based image indexing and so on not all of them are efficient in most cases. These are not efficient in terms of time and space efficiency.


Introduction part mainly deals with thefollowing:
1) Basics of image retrieval.
2) Different types of image retrieval techniques.
3) Color space models.

A. Image retrieval :

An image retrieval system is a computer system for browsing, searching and retrieving images from a large database of digital images. Most traditional and common methods of image retrieval utilize some method of adding metadata such as captioning, keywords, or descriptions to the images so thatretrieval can be performed over the annotation words.Manual image annotation is time-consuming,laborious and expensive; to address this, there hasbeen a large amount of research done on automaticimage annotation. Additionally, the increase in socialweb applications and the semantic web have inspiredthe development of several web-based imageannotation tools.

In image retrieval system features are extracted froma given query image which are then used forcomparing with the features of other images stored indatabase. In the proposed signature technique animage’s signature bi string referred as signature is anabstract representation of the color distribution of animage by bit stings of a predetermined size. Toprocess a query all the image signatures arecompared against the signature of query image usinga well defined similarity metric. The candidateimages are then retrieved and ranked according totheir similarity with query image.

B. There are mainly 3 visual attributesthrough which image retrieval happens:
Color is an extensively utilized visual attribute inimage retrieval, that often simplifies objectidentification and extraction. Color is especiallyconvenient,becauseitprovidesmultiplemeasurements at a single pixel of the image.
A Global Color Histogram (GCH) is the mosttraditional way of describing the color property of animage. The GCH for an image is constructed bycomputing the normalized percentage of color pixelsin an image corresponding to each color element.Assuming an n color model, a GCH is then an n-dimensional feature vector (h1,h2,, where hjrepresents the (usually) normalized percentage ofcolor pixels in an imagecorresponding to each color element cj . Moreformally, we can define the component hj as a uniquecombination of the values – red, green, and blue(obtained after normalization).
The visual attribute texture is a representation of thesurface of an image object. Intuitively, the term refersto properties such as smoothness, coarseness, andregularity of an image object.Shape refers to the characteristic contour of an objectthat identifies it in a meaningful form. Traditionally,shapes are represented through a set of features suchas area, axis-orientation, certain characteristic pointsetc. Shape representations are broadly divided intotwo categories, boundary-based and region-based,and the most successful representatives of these arethe Fourier Descriptor and Moment Invariantsrespectively. Retrieval of a subset of images thatsatisfy certain constraints is a central problem inshape retrieval, with the degree of similarity betweentwo images calculated as the distance betweencorresponding points.A common similarity metric is based on theEuclidean distance between the abstracted featurevectors that represent two images defined as:d(Q,I)= squareroot (summation(hj of query- hj ofimage)square))
where Q and I represents the query image and theimage in image set.

C.Color Space Model
Color space models define colors in threedimensions, such that each color is represented bythree coordinates. The difference between any twocolors is approximated to be the Euclidean distancein a uniform color space.
RGB color space model is the most commonly usedmodel, which is composed of three primary colorsred, green, blue. The primary colors are additive, thatis by varying their combinations, other colors can beobtained. The model is visualized as a unit cube withcorners of black, white, the three primary colors(red, blue, green), and the three secondary colors(cyan,magenta, yellow).
The CMY color model is based on the secondarycolors of the RGB color space model, that is – cyan(green plus blue), magenta (red plus blue) and yellow(red plus green). The subset of the Cartesiancoordinate system for the CMY color model issimilar to that of the RGB color space, except that thewhite color occupies the origin.

Author: Ravi Bandakkanavar

A Techie, Blogger, Web Designer, Programmer by passion who aspires to learn new Technologies every day. A founder of Krazytech. It's been 10+ years since I am publishing articles and enjoying every bit of it. I want to share the knowledge and build a great community with people like you.

Suggested articles for you:

2 thoughts on “Image retrieval using compact Binary signatures

Did it help? Would you like to express?