Fuzzy Techniques for Content-Based Image Retrieval

Fuzzy Techniques for Content-Based Image Retrieval

Rose Bindu Joseph P. (VIT University, India) and Ezhilmaran Devarasan (VIT University, India)
Copyright: © 2018 |Pages: 24
DOI: 10.4018/978-1-5225-5775-3.ch003

Abstract

Content-based image retrieval aims to acquire images from huge databases by analyzing their visual features like color, texture, shape, and spatial relationship. The search for superior accuracy in image retrieval has resulted in concentrating more on semantic gap reduction between the low-level features and high level human reasoning. Fuzzy theory is a prevailing methodology which helps in attaining this goal by using attributes and interpretations similar to human reasoning. The vagueness and impreciseness in image data and the retrieval process can be modeled by fuzzy sets. This chapter analyses fuzzy theoretic approaches in various stages of content-based image retrieval system. Various fuzzy-based feature descriptors are discussed along with different fuzzy classification and indexing algorithms for content-based image retrieval. This chapter also presents an overview of various fuzzy distance and similarity measures for image retrieval. A novel fuzzy theoretic retrieval for finger vein biometric images is also proposed in this chapter with experiment and analysis.
Chapter Preview
Top

Introduction

During the past few decades, volume of digital image databases has grown exponentially because of the rapid advancement of internet, modern cameras and technologies. The quick surge of image data has amplified efficient accessibility in various fields such as security, medical image archives, multimedia encyclopedia, geographical information systems and biometric databases. In view of this, many content-based image retrieval (CBIR) systems have been established by utilizing color descriptors, texture or shape features for recovering preferred images from an assortment.

Even though many disadvantages of the outdated text-based retrieval systems have been overcome by CBIR systems, only low-level features are extracted by them to denote image data. The basic nature of subjective and fuzzy understanding and intuition of individuals cannot be handled by these rigid low-level measures. Also, for a visual content, different inferences and descriptions would be given by different individuals. In practical situations, high level reasoning and features are used by individuals for interpreting and recalling similar images and measuring the similarities between them. The need for improving the accuracy in retrieval process has forced to shift the focus of research in image retrieval into semantic gap reduction that lies among lower grade visual structures and competent semantics possessed by human. Fuzzy theory is a powerful tool to realize this goal which uses attributes and inferences similar to human reasoning. Furthermore, image data, specifications for image query and measures of similarity in retrieval problems comprise of fuzziness and imprecision. The concepts modeled by fuzzy set theory have no exact boundary between membership and non-membership and the change is gradual rather than abrupt. Fuzzy approaches can be used to model the vagueness and to represent and process imprecise and uncertain data. Fuzzy techniques provide a data model which is intuitive and user friendly that can take into account subjectivity and uncertainties. Human thinking and decision-making skills can be emulated in machines using the powerful reasoning algorithms of fuzzy theory.

Different researchers have proposed different fuzzy feature descriptors by replacing low level visual features for retrieving images similar to query image. Fuzzy attributed relational graphs are fuzzy image descriptors that can robustly represent objects in images along with their attributes and spatial relations. Another fuzzy descriptor that is used in CBIR is Fuzzy color histogram (FCH) which been proposed as a competent feature descriptor that can overcome the drawbacks of traditional non-fuzzy descriptors efficiently.

Traditional techniques for image classification classify images mainly into discrete classes. A pixel is allocated into a class having highest degree of similarity. But it is not preferable in practical problems to allocate membership of image data completely into one class. Representation of images into continuous classes is the best alternative approach to this discrete representation. This can be effectively achieved by assigning a membership degree indicating the relative strength of inclusion into one particular class. A fuzzy support vector machine (FSVM) is an efficient alternative to classical SVM which reduces the effects of outliers and noises by using an appropriate fuzzy membership function. Fuzzy SVM performs powerfully even when the data is small in size. Fuzzy inference system is another fuzzy classifier widely used in image classification and retrieval which uses a series of appropriate fuzzy inference commands to classify images into fuzzy classes. Fuzzy c-means clustering (FCM) is a prevalent algorithm that works with the concepts of fuzzy theory. In this, each image feature is assigned a fuzzy membership of belongingness to a cluster.

Fuzzy theory offers a wide range of distance measures as well as similarity functions such as fuzzy Minkowsky distance, fuzzy Euclidean distance and fuzzy set theoretic similarity measures. The extracted visual features are transformed into fuzzy plane before calculating the appropriated fuzzy distance or similarity measures between them.

Complete Chapter List

Search this Book:
Reset