NeuroMerge: an approach for merging heterogeneous features in content-based image retrieval systems

Visual database systems require efficient and effective mechanisms for content-based retrieval. Content of an image can be expressed in terms of different features such as shape, color, texture, or text annotations. Retrieval based on each of these individual features can result in a different set of images. Thus, there is a need to merge the results obtained from these individual heterogeneous features. Most of the existing approaches assume a linear relationship between different features, and also require the user to directly assign weights to features. We propose NeuroMerge, a neural network based model, to merge the results from heterogeneous features. Using a set of training data, NeuroMerge assigns weights to the features and removes the burden from users. This model can be used to determine the nonlinear relationship between features so that more accurate similarity comparison between images can be supported. Experimental results are presented to demonstrate efficiency and high accuracy of this method.