*********************************
There is now a CONTENT FREEZE for Mercury while we switch to a new platform. It began on Friday, March 10 at 6pm and will end on Wednesday, March 15 at noon. No new content can be created during this time, but all material in the system as of the beginning of the freeze will be migrated to the new platform, including users and groups. Functionally the new site is identical to the old one. webteam@gatech.edu
*********************************
Title: Weakly Supervised Semantic Labeling of Migrated Seismic Data
Committee:
Dr. Ghassan AlRegib, ECE, Chair , Advisor
Dr. James McClellan, ECE
Dr. Mark Davenport, ECE
Dr. Ying Zhang, ECE
Dr. Zhigang Peng, EAS
Abstract:
Deep learning has revolutionized the fields of machine learning and computer vision. However, the availability of annotated data to train state-of-the-art deep networks is one of the main bottlenecks to the successful application of deep learning, especially to applications like seismic interpretation where annotated data is extremely limited. In this thesis, we develop a weakly-supervised framework for the semantic labeling of large seismic volumes. This framework involves developing a state-of-the-art texture similarity measure and using it for retrieving large numbers of images with high visual similarity to exemplar images for each target class. Images with high visual similarity can be assigned image-level labels matching those of the exemplar images used to retrieve them. A novel weakly-supervised label mapping algorithm, based on orthogonal non-negative matrix factorization, is then used to transform these image-level labels into pixel-level labels that encode the locations of the target classes within each image. Finally, these weak pixel-level labels are used to train deep convolutional networks for the semantic labeling of various seismic structures and lithostratigraphic units within large seismic volumes. A special loss function is introduced to help the networks learn effectively when trained with weak labels. The benefit of this work is that it enables the training and deployment of deep learning models to new application domains-- such as seismic interpretation-- where sufficient quantities of labeled data are not available, and annotation costs are prohibitively expensive.
