The choice for low level representation of images makes a significant difference in image retrieval performance regardless whether the retrieval engine works on low level or high level features. The rich diversity of natural images makes this decision crucial. This paper reports some initial result of our research in image semantics. The main objective of this part is to find an effective region based image representation that has highly potential clue for image semantics. Therefore, this paper compares retrieval effectiveness of different region based representations for images. Our study shows that complex techniques like automatic segmentation based methods do not guarantee improved image retrieval performance compared to simple approaches. This paper critically evaluates the merits of these approaches.
展开▼