Date of Graduation
Statler College of Engineering and Mineral Resources
Industrial and Managements Systems Engineering
Measuring visual similarity between two images is useful in several multimedia applications such as visual search and image retrieval. However, measuring visual similarity between two images is an ill-posed problem which makes it a challenging task.This problem has been tackled extensively by the computer vision and machine learning communities. Nevertheless, with the recent advancements in deep learning, it is now possible to design novel image representations that allow systems to measure visual similarity more accurately than existing and widely adopted approaches, such as Fisher vectors. Unfortunately, deep-learning-based visual similarity approaches typically require post-processing stages that can be computationally expensive. To alleviate this issue, this thesis describes deep-learning-based visual image representations that allow a system to measure visual similarity without requiring post-processing stages. Specifically, this thesis describes max-pooling-based aggregation layers that combined with a convolutional-neural-network-based produce rich image representations for image retrieval without requiring an expensive post-processing stages. Moreover, the proposed max-pooling-based aggregation layers are general and can be seamlessly integrated with any existing and pre-trained networks. The experiments on large-scale image retrieval datasets confirm that the introduced image representations yield visual similarity measures that achieve a comparable or better retrieval performance than state-of-the art approaches, without requiring expensive post-processing operations.
Kazkaz, Mohammad Eiad, "Aggregating Deep Features For Image Retrieval" (2017). Graduate Theses, Dissertations, and Problem Reports. 5951.