Deep Learning based Image Retrieval with Weight Learning and Diffusion Modelling

Image retrieval is to retrieve or search digital images from large databases. The difficulty of image retrieval lies at is the semantic gap, and that the images in a database usually do not have any labels. The research of image retrieval based on deep learning aims to design and propose new ideas, methods, and algorithms to improve its performance. From the perspective that image retrieval is generally an unsupervised task, this thesis firstly analyses some typical unsupervised deep learning models. Under the unsupervised settings, this work compares these deep learning models with the classical BoW model. This gains valuable insights into unsupervised learning and network training for the subsequent image retrieval studies in this thesis.