Multiresolution wavelet transform and supervised learning for content-based image retrieval

We focus on the definition of an effective strategy that allows the user to pose a visual query and retrieve a set of images from a database that satisfy his criteria of pictorial similarity without requiring any semantic expression of them. The strategy exploits a multiresolution wavelet transform to effectively describe image content. The salient features of the images are coded in signatures of predefined lengths which are compared in the retrieval phase by applying a similarity measure the system has pre-learned, using a regression model for ordinal responses, from a learning set of "very similar", "rather-similar", "not-very-similar", and "different" pairs of images. Some experimental results demonstrating the effectiveness of this approach are reported.