A Novel Multile Salient Regions Combination-Based Image Retrieval Scheme With Color And Texture Features
Mingxin Zhang,Zhaowei Shang,Wenhui Rijan,Junyi Shen
DOI: https://doi.org/10.1109/ICWAPR.2008.4635758
2008-01-01
Abstract:In current common salient regions-based image retrieval schemes, salient regions are individually disposed, and one could not retrieve the objects with distinct indispensable regions. Herein, we propose one multiple salient regions combined (MSRC) image retrieval scheme, which uses all the salient regions of target object images to search the implicit objects in querying images for their counterparts. If and only if all the salient regions that represent the target objects have their counterparts, the querying image could be retrieved. Meanwhile, the current saliency metrics fail to reflect the saliency of implicit regions, and we design a novel saliency metric in terms of dominant colors and texture characteristics, which have crucial roles to distinguish different regions or semantic objects. Finally, with the proposed saliency metric rather than the wavelet transform modulus maxima (WTMM) point density and scale saliency, one can extract distinct salient regions consistent with human semantic perceptions. Experiment Results indicated that the proposed MSRC scheme outbalances the common salient regions-based scheme under different saliency metrics. Its average retrieval accuracy can reach to 0.84, while the average accuracies for scale saliency and WTMM point density are only 0.75 and 0.6425 respectively. An accuracy increase 0.20 over that of WTMM point density can be achieved by the proposed saliency metric.