Landmark Image Searching with Inattentive Salient Regions

Manual tagging has an important impact to performance of image/video searching by keyword. However, users usually mark tags only landmarks are as on only a few images in library and leave most contents untagged. If landmarks from different places are look alike, it is hard to distinguish even though surroundings are totally different. Rather than using only manual tags of highlight landmark, we proposed to use automatic tags of distinct inattentive salient regions to improve the search accuracy. Inattentive salient regions are unimportant areas to the users but highly relevant to the landmark. We determine salient regions by SIFT descriptors, find regions, find inattentive regions, and represent the relationships between inattentive regions and landmarks as an extra index. Dataset in the experiment is composed of 2,917 images of various landmark locations from public databases. The experimental results demonstrate 10% improvement of accuracy between using highlight landmark only and applying our proposed method.