Statistical Approach for Image Retrieval

• Main Author: Khor, Siak Wang
• Format: Thesis
• Language: English; English
• Published: 2007
• Online Access: http://psasir.upm.edu.my/id/eprint/441/1/1600481.pdf; http://psasir.upm.edu.my/id/eprint/441/
• Institution: Universiti Putra Malaysia

Since the emergence of Internet, a gigantic volume of images have been uploaded into the Internet from time to time. Relying on the traditional text-based search approach to locate the required images could no longer meet the diverse needs of users. This persistent trend has demanded a more sophisticated search algorithm on these images. One of the popular and common approaches for image search is Content-based Image Retrieval or CBIR for short, i.e. retrieval of images based on their visual contents such as shapes, colours, textures etc. Of all the visual contents identifiable from an image, colour is considered to be the commonest visual attribute that aids in image retrieval. Works on colour-based image retrieval systems are largely based on the use of colour histogram, which has been noted to suffer from a major drawback, i.e. absence of spatial information, which is also an important requirement for an accurate retrieval result. In this thesis, a novel method based on the modified generic framework of CBIR is proposed. This technique, formally known as Image Retrieval Using Statistical-based Approach is based on the idea of grouping pixels with similar colour codes within an image. From these grouped pixels, they are sorted in descending order of pixel count, which intuitively identifies dominant colours within an image. Statistical information, i.e. means and standard deviations will then be derived from these sorted groups. The extracted statistical information will be stored in both text files and matrixes, which will be used to aid in the image retrieval process. The system has also included some adjustable parameters, such as window size, CC percentage similarity, which can be used to improve retrieval accuracy. This statistical-based approach has been tested on the standard UCID image collection where it has shown improved results, with an average precision value of about 70% as compared to an approximate value of 25% using the histogram-based approach, in term of retrieval accuracy.