Investigating the Behavior of Compact Composite Descriptors in Early Fusion, Late Fusion and Distributed Image Retrieval

Abstract

In Content-Based Image Retrieval (CBIR) systems, the visual content of the images is mapped into a new space named the feature space. The features that are chosen must be discriminative and sufficient for the description of the objects. The key to attaining a successful retrieval system is to choose the right features that represent the images as unique as possible. A feature is a set of characteristics of the image, such as color, texture, and shape. In addition, a feature can be enriched with information about the spatial distribution of the characteristic that it describes. Evaluation of the performance of low-level features is usually done on homogenous benchmarking databases with a limited number of images. In real-world image retrieval systems, databases have a much larger scale and may be heterogeneous. This paper investigates the behavior of Compact Composite Descriptors (CCDs) on heterogeneous databases of a larger scale. Early and late fusion techniques are tested and their performance in distributed image retrieval is calculated. This study demonstrates that, even if it is not possible to overcome the semantic gap in image retrieval by feature similarity, it is still possible to increase the retrieval effectiveness.