Fusion of region based extracted features for instance- and class-based CBIR applications

Abstract

One of the major challenges in content-based image retrieval (CBIR) is understanding the actual semantic meaning of images. Generally, natural images possess a rich collection of different visual features, each of which have different contributions to the image semantics. These vary drastically in their local sub-regions. However, most of the state-of-the-art CBIR techniques give equal importance to all the visual features, which are not always effective for generating relevant retrieval outcomes. To overcome these issues, a CBIR scheme has been proposed which considers the image semantic properties of the different regions with their positional significance in the overall image. In this paper, a novel regions-of-attention (ROA) based feature extraction and fusion scheme for efficient image retrieval is suggested. Initially, a salient key-points and opponent-color feature based scheme is used to locate ROA of the image. Multi-directional texture, and spatial correlation-based color features are extracted from this ROA of the image to obtain most of the image’s semantic meanings. Spatial features from salient non-ROA local regions are also obtained to capture vital background information. A dissimilarity factor-based fusion scheme then is utilized to fuse the ROA and non-ROA features effectively, considering the importance and uniqueness of each feature to support the positional invariance notion in their respective sub-regions. Finally, the fused feature vector is used to perform instance- and class-based image retrieval in five different image datasets using different classifiers. The outcomes of the retrieval experiment have shown competitive performances as compared to the current state-of-the-art architectures.