When no explicit annotation exists, web-based image search engines cannot understand the actual content of images. As a result they return several noisy data in most searches, leading to low precision and recall. Furthermore, usually a large number of images are returned in several sequential pages, a visualization way that is not view-and transmission-efficient. To overcome these problems an automatic wrapper-based image retrieval and presentation system is proposed that performs significantly better than other search engines. In particular our novel system exploits the format of multimedia sharing web sites to discover the underlying structure in order to finally infer and extract multimedia files and corresponding associated keywords from the web pages. Afterwards the gathered content is properly organized in a multiscale tree structure for efficient visualization. By this way, users can select only paths-of-interest in the tree structure, thus detecting the needed content much faster and spending less bandwidth compared to current solutions. Experimental results are presented that indicate the interesting performance of the proposed architecture.