Date of Award
CBIR, Dimensionality reduction, Feature descriptor, Quadtree representation, Siamese networks, Similarity detection, Content-based image retrieval
Due to modern technological advancements, the pervasiveness and complexity of images have remarkably increased. Searching databases for similar visual content, i.e., Content-Based Image Retrieval (CBIR), remains an open research problem. In this thesis, we propose a novel CBIR approach, in which each symbolic image has a quadtree representation consisting of SIFT-based orientational keypoints. Every quadrant node in the tree represents the dominant orientation of a region in the image. The quadtree image representation is used for bitwise signature indexing and image similarity measurement. Also, we convert each quadtree image representation to a trainable feature vector for use in the K-Nearest Neighbour algorithm and Siamese Deep Neural Networks. The proposed approaches are evaluated using mean average precision (mAP), precision, recall, f-score and contrastive loss on three different image datasets. Our results indicate that, for complex images, orientational quadtrees are significantly more accurate than spatial quadtrees. Further, the derived feature vectors can be used in other machine learning or deep learning methods for training, ensemble, boosting, aggregation or embedding.
Adil, Eisa, "Content-Based Image Retrieval using Hierarchical Decomposition of Feature Descriptors" (2021). Electronic Theses and Dissertations. 8779.