Image classification and segmentation are two fundamental tasks in the field of computer vision. Deep fully supervised learning techniques have recently achieved remarkable performance on these two tasks, by leveraging a large amount of labelled training data. However, there are scarce image-level labels in many real-world applications for image classification, let alone pixel-level dense annotations for image segmentation. This thesis focuses on two practical tasks, i.e., (i) coral image classification, for which algorithms based on transfer learning and data augmentation techniques are proposed; (ii) semantic segmentation, for which weakly supervised techniques based on multi-scale learning and auxiliary learning are proposed using only image-level labels.
|Qualification||Doctor of Philosophy|
|Award date||7 Jun 2021|
|Publication status||Unpublished - 2021|