A Feature Level Fusion in Similarity Matching to Content-Based Image Retrieval

This paper presents a fusion-based similarity matching framework for content-based image retrieval on a combination of global, semi-global and local region specific features at different levels of abstraction. In this framework, an image is represented by global color and edge histogram descriptors, semi-global color and texture descriptors from grid based overlapping sub-images and local color features from a clustering-based segmented regions. As a result, image similarities are obtained through a weighted combination of overall similarity fusing global, semi-global and local region-based image level similarities. This fusing approach decreases the impact of inaccurate segmentation and increases retrieval effectiveness as constituent features are of a complementary nature. The experimental results on a general-purpose image database indicate that the aggregation or fusion-based technique provides an effective and flexible tool for similarity calculation based on a combination of descriptors from different levels of image representation