Content-based image retrieval (CBIR) considers the characteristics of the image itself, for example its shapes, colors and textures. The current approaches to CBIR differ in terms of which image features are extracted. Recent work deals with combination of distances or scores from different and independent representations. This work attempts to induce high level semantics from the low level descriptors of the images. In this paper, we propose a new approach that integrates techniques of salient, color and texture features. Our approach extracts interest salient regions that work as local descriptors. A greedy graph matching algorithm with a proposed modified scoring function is applied to determine the final image rank. The proposed approach is appropriate for accurately retrieving images even in distortion cases such as geometric deformations and noise. This approach was tested on proprietary image databases. Also an offline case study is developed where our approach is tested on images retrieved from Google keyword based image search engine. The results show that a combination of our approach as a local image descriptor with another global descriptor outperforms other approaches. |
