Towards Better Image Embeddings Using Neural Networks

Abstract

The primary focus of this dissertation is to study image embeddings extracted by neural networks. Deep Learning (DL) is preferred over traditional Machine Learning (ML) for the reason that feature representations can be automatically constructed from data without human involvement. On account of the effectiveness of deep features, the last decade has witnessed unprecedented advances in Computer Vision (CV), and more real-world applications are expected to be introduced in the coming years.

A diverse collection of studies has been included, covering areas such as person re-identification, vehicle attribute recognition, neural image compression, clustering and unsupervised anomaly detection. More specifically, three aspects of feature representations have been thoroughly analyzed. Firstly, features should be distinctive, i.e., features of samples from distinct categories ought to differ significantly. Extracting distinctive features is essential for image retrieval systems, in which an algorithm finds the gallery sample that is closest to a query sample. Secondly, features should be privacy-preserving, i.e., inferring sensitive information from features must be infeasible. With the widespread adoption of Machine Learning as a Service (MLaaS), utilizing privacy-preserving features prevents privacy violations even if the server has been compromised. Thirdly, features should be compressible, i.e., compact features are preferable as they require less storage space. Obtaining compressible features plays a vital role in data compression.

Towards the goal of deriving distinctive, privacy-preserving and compressible feature representations, research articles included in this dissertation reveal different approaches to improving image embeddings learned by neural networks. This topic remains a fundamental challenge in Machine Learning, and further research is needed to gain a deeper understanding.