Multispectral Image Retrieval Using a Distance Based on Vector Quantization

摘要

This paper proposes a new method of similar image retrieval based on the spectral distribution of a multispectral image. The greatest advantage of multispectral imaging is that the color and reflectivity of the object can be determined accurately. In similar image retrieval these physical variables must be discriminated and accurately evaluated. However, in the conventional retrieval method, the color composition ratio is emphasized rather than evaluation of the color (as a physical variable). In the histogram intersection method, for example, which is a form of similarity calculation based on the histogram, the similarity is calculated from only the shape of the histogram, that is, the color composition ratio. Thus, the similarity may be understated simply because the composition ratio is different, which makes this approach unsuitable for multispectral image retrieval, in which the color similarity itself is to be evaluated. This paper proposes a method of retrieval which does not depend on the color composition ratio (a physical variable), but is based on the similarity of the physical variable itself. The method is based on the distance between images, which accurately represents the difference in the physical variables by vector quantization. The effectiveness of the proposed method is demonstrated by an experiment on similar image retrieval for multispectral images.