Use this URL to cite or link to this record in EThOS:
Title: Scalable image retrieval based on hand drawn sketches and their semantic information
Author: Bozas, Konstantinos
ISNI:       0000 0004 5355 2490
Awarding Body: Queen Mary University of London
Current Institution: Queen Mary, University of London
Date of Award: 2014
Availability of Full Text:
Access from EThOS:
Access from Institution:
The research presented in this thesis aims to extend the capabilities of traditional content-based image retrieval systems, towards more expressive and scalable interactions. The study focuses on machine sketch understanding and its applications. In particular, sketch based image retrieval (SBIR), a form of image search where the query is a user drawn picture (sketch), and freehand sketch recognition. SBIR provides a platform for the user to express image search queries that otherwise would be di cult to describe with text. The research builds upon two main axes: extension of the state-of-the art and scalability. Three novel approaches for sketch recognition and retrieval are presented. Notably, a patch hashing algorithm for scalable SBIR is introduced, along with a manifold learning technique for sketch recognition and a horizontal ip-invariant sketch matching method to further enhance recognition accuracy. The patch hashing algorithm extracts several overlapping patches of an image. Similarities between a hand drawn sketch and the images in a database are ranked through a voting process where patches with similar shape and structure con guration arbitrate for the result. Patch similarity is e ciently estimated with a hashing algorithm. A spatially aware index structure built on the hashing keys ensures the scalability of the scheme and allows for real time re-ranking upon query updates. Sketch recognition is achieved through a discriminant manifold learning method named Discriminant Pairwise Local Embeddings (DPLE). DPLE is a supervised dimensionality reduction technique that generates structure preserving discriminant subspaces. This objective is achieved through a convex optimization formulation where Euclidean distances between data pairs that belong to the same class are minimized, while those of pairs belonging to di erent classes are maximized. A scalable one-to-one sketch matching technique invariant to horizontal mirror re ections further improves recognition accuracy without high computational cost. The matching is based on structured feature correspondences and produces a dissimilarity score between two sketches. Extensive experimental evaluation of our methods demonstrates the improvements over the state-of-the-art in SBIR and sketch recognition.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available
Keywords: Electronic Engineering ; Computer Science