Learning-based Image Compression for Unified Human and Machine Vision
This thesis explores new ways to compress images for both human vision and machine learning algorithms. Targeting compression for human vision, we provide a new deep learning-based method that compresses images more efficiently than current methods. We develop a way to adapt pre-trained deep learning-based compressors to new data over time to improve their flexibility. Targeting compression for machine vision, we explore deep neural network feature compression in various scenarios. We develop an efficient feature compression method for edge-cloud systems. We extend it to a flexible compression system that can adjust the three-way trade-off between rate, prediction accuracy, and neural network model complexity. Finally, we propose a unified compression system that integrates image compression for both human and machine vision. Our system enables efficient compressed-domain inference of vision tasks such as classification and segmentation while retaining the ability to reconstruct the image using the compressed feature.
We believe this thesis contributes to advancing the lossy data compression research.
History
Degree Type
- Doctor of Philosophy
Department
- Electrical and Computer Engineering
Campus location
- West Lafayette