Hessian-Laplace feature detector and Haar descriptor for image matching

Abstract

In recent years feature matching using invariant features has gained significant importance due to its application in various recognition problems. Such techniques have enabled us to match images irrespective of various geometric and photometric transformations between images. The thesis being presented here focusses on developing such a feature matching technique which can be used to identify corresponding regions in images. A feature detection approach is proposed, which finds features that are invariant to image rotation and scaling, and are also robust to illumination changes. A description is computed for each feature using the local neighborhood around it and then acts as a unique identifier for the feature. These feature identifiers (or feature descriptors) are then used to identify point to point correspondences between images. A systematic comparison is made between this feature detector, and others that are described in the literature. Later in this work, we apply the feature matching technique developed here to perform image retrieval for panoramic images. Our objective here is to retrieve a panoramic image similar to a query image from a database. We show how such a retrieval task can be performed by giving results for both indoor and outdoor sequences.