Robust Hierarchical Indexing based on Texture Features

In this paper, we present a hierarchical indexing method based on texture characterization for image retrieval. The novelty of our contribution is the hierarchical structure of the index: it exploits the multiresolution formulation of Wavelet Transforms to define a new set of approximated versions of the images for each level of resolution. On this set, the algorithm extracts significant signatures by means of statistical correlations; the experimental results and the analysis of computational complexity have proved that the algorithm presents the best performance at the highest level of the indexing hierarchy, where the computational complexity is the lowest. Our method has been evaluated by the following methodologies: (a) the study of the computational complexity for signature generation; (b) the comparison with analogous methods based on texture analysis by reporting the performance obtained on the same database (Brodatz); and (c) the evaluation of the robustness of the hierarchical indexing in different color spaces, by querying the database with different versions of the original images obtained by noise addition (gaussian and scanner acquisition noise and lossy compression distortion), brightness and contrast enhancement, color and scale adjustment and rotation. Even if our method is designed for texture databases, experiments show satisfactory results also on a real heterogeneous photographic database. This confirms the possibility of exploiting our method as a low computational complexity indexing tool based on texture characterization in a broader system for hierarchical content-based retrieval.