Multi evidence fusion scheme for content-based image retrieval by clustering localised colour and texture features

Al-Jubouri, Hanan (2015) Multi evidence fusion scheme for content-based image retrieval by clustering localised colour and texture features. Doctoral thesis, University of Buckingham.

Preview

Text Al-Jubouri_2015_Multi.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (9MB) | Preview

Abstract

Content-Based Image Retrieval (CBIR) is an automatic process of retrieving images according to their visual content. Research in this field mainly follows two directions. The first is concerned with the effectiveness in describing the visual content of images (i.e. features) by a technique that lead to discern similar and dissimilar images, and ultimately the retrieval of the most relevant images to the query image. The second direction focuses on retrieval efficiency by deploying efficient structures in organising images by their features in the database to narrow down the search space. The emphasis of this research is mainly on the effectiveness rather than the efficiency. There are two types of visual content features. The global feature represents the entire image by a single vector, and hence retrieval by using the global feature is more efficient but often less accurate. On the other hand, the local feature represents the image by a set of vectors, capturing localised visual variations in different parts of an image, promising better results particularly for images with complicated scenes. The first main purpose of this thesis is to study different types of local features. We explore a range of different types of local features from both frequency and spatial domains. Because of the large number of local features generated from an image, clustering methods are used for quantizing and summarising the feature vectors into segments from which a representation of the visual content of the entire image is derived. Since each clustering method has a different way of working and requires settings of different input parameters (e.g. number of clusters), preparations of input data (i.e. normalized or not) and choice of similarity measures, varied performance outcomes by different clustering methods in segmenting the local features are anticipated. We therefore also intend to study and analyse one commonly used clustering algorithm from each of the four main categories of clustering methods, i.e. K-means (partition-based), EM/GMM (model-based), Normalized Laplacian Spectral (graph-based), and Mean Shift (density-based). These algorithms were investigated in two scenarios when the number of clusters is either fixed or adaptively determined. Performances of the clustering algorithms in terms of image classification and retrieval are evaluated using three publically available image databases. The evaluations have revealed that a local DCT colour-texture feature was overall the best due to its robust integration of colour and texture information. In addition, our investigation into the behaviour of different clustering algorithms has shown that each algorithm had its own strengths and limitations in segmenting local features that affect the performance of image retrieval due to variations in visual colour and texture of the images. There is no algorithm that can outperform the others using either an adaptively determined or big fixed number of clusters. The second focus of this research is to investigate how to combine the positive effects of various local features obtained from different clustering algorithms in a fusion scheme aiming to bring about improved retrieval results over those by using a single clustering algorithm. The proposed fusion scheme integrates effectively the information from different sources, increasing the overall accuracy of retrieval. The proposed multi-evidence fusion scheme regards scores of image retrieval that are obtained from normalizing distances of applying different clustering algorithms to different types of local features as evidence and was presented in three forms: 1) evidence fusion using fixed weights (MEFS) where the weights were determined empirically and fixed a prior; 2) evidence fusion based on adaptive weights (AMEFS) where the fusion weights were adaptively determined using linear regression; 3) evidence fusion using a linear combination (Comb SUM) without weighting the evidences. Overall, all three versions of the multi-evidence fusion scheme have proved the ability to enhance the accuracy of image retrieval by increasing the number of relevant images in the ranked list. However, the improvement varied across different feature-clustering combinations (i.e. image representation) and the image databases used for the evaluation. This thesis presents an automatic method of image retrieval that can deal with natural world scenes by applying different clustering algorithms to different local features. The method achieves good accuracies of 85% at Top 5 and 80% at Top 10 over the WANG database, which are better when compared to a number of other well-known solutions in the literature. At the same time, the knowledge gained from this research, such as the effects of different types of local features and clustering methods on the retrieval results, enriches the understanding of the field and can be beneficial for the CBIR community.

Actions (login required)

View Item