Content-Based Image Retrieval System By Multi-Dimensional Feature Vectors
Shinya Asami, Hiroaki Kotera · Technical programs and proceedings/Technical program and proceedings · 2003
This paper introduces our approach to content-based image retrieval by means of key image in place of keywords. The color, spatial frequency, and shape features are extracted from the image sources and spanned in multi-dimensional feature vector space. The mutual color distances between segmented key color areas are compared to narrow the candidates for key image with desired color-tone. The spatial frequency features are represented by the limited DCT components, where low-to-middle frequency spectra are selected in circular zonal sectors. Here the mutual correlations were taken for only 7 DCT components to discriminate the differences in image textures. In addition, the compositional or shape features are simply extracted by down sampling and labeling process. The correlation between the mosaic bi-level patterns after down sampling was very useful to find the rough similarity in image structure. The paper presents experiments on the stamp or facial image retrieval.