Degree Name

Doctor of Philosophy


School of Computing and Information Technology


Instance image retrieval aims at retrieving all the images containing the same instance that is depicted in the query image provided by the users from an image gallery. During the past decades, it has attracted the attention of a lot of researchers in the computer vision area and has been successfully applied to a variety of realistic applications in the present world. However, as one of the unsupervised learning tasks, the datasets for instance image retrieval tasks are usually provided with no human annotations. This makes it challenging to neither learn a reliable feature extraction model nor build precise similarity measure to conduct retrieval. Therefore, this thesis focuses on mining more useful information within the retrieval dataset to help achieve better retrieval performance. In particular, this thesis unfolds its investigation on instance image retrieval from the following three aspects.

Firstly, this thesis argues that, even with the powerful deep features, uniformly apply Euclidean distance to measure the similarity between the gallery images and the query images is not precise enough. To address this issue, this thesis proposes a framework to construct a query-adaptive similarity measure that captures the discriminative characteris-tic of the given query without online training. In particular, it identifies the characteristic of a query by aggregating the unique characteristics of proper images in a dataset. This not only helps to build a better similarity measure to enhance retrieval but also switches classifier training from online to offline to shorten system response time. Compared with the existing state-of-the-art query expansion methods, our method outperforms them on most of the instance retrieval datasets. The experiment conducted on a diffusion process based image retrieval task also demonstrates the advantage of the proposed method.

FoR codes (2008)




Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.