Gustav Hanning presents his Master's Thesis: Image Retrieval Re-ranking using Graph Neural Networks, Tuesday 7/5 at 10:15 in MH:309A.

Abstract:
Image retrieval is the task of finding images in a database similar to a given query image. The retrieved images, typically a small subset of the entire database, are initially ordered based on their similarity with the query. They can subsequently be re-ranked to improve the retrieval accuracy. Database images that are relevant to the query should increase in rank and vice versa.

In this thesis the re-ranking process is modeled as a graph neural network. The nodes of the graph are the query and retrieved database images. For each node an affinity vector is computed which encodes the visual similarity between the image and a set of anchor images. The vectors are refined by message passing between nodes, using self-attention. Database images are re-ranked according to the similarity between their refined affinity vector and that of the query.

The network is trained on a large-scale dataset and evaluated against three other re-ranking algorithms. Results show that the method proposed in the thesis achieves significantly higher precision.

Supervisors:
Viktor Larsson, Centre for Mathematical Sciences
Gabrielle Flood, Centre for Mathematical Sciences

Examiner:
Magnus Oskarsson, Centre for Mathematical Sciences