Inverse scale invariant feature transform models for object recognition and image tagging.

This thesis presents three novel image models based on Scale Invariant Feature Transform (SIFT) features and the k-Nearest Neighbors (k-NN) machine learning methodology. While SIFT features characterize an image with distinctive keypoints, the k-NN filters away and normalizes the keypoints with a two-fold goal: (i) compressing the image size, and (ii) reducing the bias that is induced by the variance of keypoint numbers among object classes. Object recognition is approached as a supervised machine learning problem, and the models have been formulated using Support Vector Machines (SVMs). These object recognition models have been tested for single and multiple object detection, and for asymmetrical rotational recognition. Finally, a hierarchical probabilistic framework with basic object classification methodology is formulated as a multi-class learning framework. This framework has been tested for automatic image annotation generation. Object recognition models were evaluated using recognition rate (rank 1) whereas the annotation task was evaluated using the well-known Information Retrieval measures: precision, recall, average precision and average recall.