Method and system for multimedia access, indexing and retrieval are traditionally based on keywords manually associated with text, videos and images. However, this activity is extremely critical and time consuming when the amount of multimedia files in the digital collections increases. For scientific community, one of the main research interests becomes the development and implementation of automatic image classification methods to understand the semantic content of image and videos. The challenge of this topic is the definition and extraction of robust low-level features to index and retrieve the multimedia files from image databases by comparing similarity distances between features. In order to solve these problems, FUB has developed an effective and rapid clustering and classification algorithm that has resulted particularly efficient in querying heterogeneous big collections of images.
Given a multimedia database, the proposed method constructs homogeneous image clusters with similar structural information and presents them to the user. The effectiveness of the approach is strictly related to an efficient and robust modelling of the images by means of low-level descriptors. This key step of the classification procedure is built in wavelet domain by accurately selecting image pattern with high structural information. Specifically, the images are locally represented by spatio-temporal descriptors computed from Laguerre-Gauss wavelet coefficients and Zernike moments. This image modelling allows simplifying the classification process by a number of comparison operations organised in a tree structure.
The final ordering of the clusters retrieved in response to the user query, is obtained by applying the Breadth First Search (BSF) algorithm to find the connected sub graphs within the binary matrix computed by the thresholding of the probability matrix related to the cluster fusion process.