微小な形状特徴に基づくテクスチャ画像の検索手法

抄録

For the database of natural texture images, a retrieval method using a texture image itself as a key image of retrieval has been developed. The feature of a texture image cannot always be clearly descried, and a retrieval method, which uses the texture image itself as its retrieval key, seems to be preferable. Based on the reason, a method taking notice of micro-shapes features in texture images has been realized by use of shape-pass filter banks proposed by H.Tamura et al. In the experiments, each of 139 sheets of texture images with 768 x 768 pixels being prepared and divided to nine sheets of equal area, one has been assigned as the key image and the other eight registered to DB for all 139 original images. Feature vectors with nine dimensions are calculated through the filter banks of micro-shapes and registered. At image retrieval, the inner products of the feature vectors of the key image and the registered DB images are calculated, and the images with the value of the inner product of nearly equal to one are selected as to be retrieved. The results of the experiments have shown the effectiveness of the method.