Image Features Based on a Mixed Fractal Model and Evaluation of Their Effectiveness in Image Retrieval

Many research works on content-based image retrieval have made use of image features. On the other hand, many kinds of image features have also been developed in the research field of wmputer vision and image analysis. The fractal signature is an example, which is based on the assumption that an image consists of a single fractal. For complex natural objects like grass and trees, however, a mixed fractal model is more suitable, which assumes an image consists of multiple different fractals. The naive use of the fractal signature for a mixed fractal images causes the problem of missing information. In this paper we propose a new image feature called the average fractal signature and its variants to solve this problem. We also demonstrate the effectiveness of the features by experiments classifying real photograph images into four object types, grass, trees, clouds, and waves. To perform the experiments in the context of image retrieval we used a simple box decomposition strategy for image partitioning without a sophisticated segmentation technique. We obtained good classification results for grass and trees by using new image features.