We propose a similarity measuring method for images based on our proposed feature extraction algorithm. The method extracts features of an image indirectly by utilizing a number of reference images. Most of proposed methods, categorized as Content-Based Image Retrieval, extract features from images by utilizing image analysis. Results of those methods do not always suit well to users' demands, since a definition of a similarity differs according to aims of retrievals or users' preferences. However, it is difficult for an user to extract features from an image in a different perspective without any special knowledge. In our method, a feature vector is calculated as a quantified value which approximates correlations in difference matrices each of which is generated from an image and one of reference images. Thus, users can easily change a feature space to represent features of images by selecting different reference images. This significant characteristic of the method is expected to be effective to achieve a similarity measurement for images based on users' demands. In this paper, we give a detailed illustration of our method and evaluate its performance through experiments.