Transform Domain Deep Neural Network Layers and Their Applications

This dissertation presents a unified research framework that integrates orthogonal transform theory with deep neural network architectures to achieve efficient data compression, representation learning, and image correction. The motivation stems from the increasing demand for accurate and resource-efficient data processing in biomedical and industrial systems, where large-scale sensor data must be transmitted and reconstructed under strict computational and bandwidth constraints. Traditional deep learning methods provide strong modeling capability but often require excessive parameters and energy consumption, while classical transformbased approaches offer interpretability and sparsity yet lack adaptability to nonlinear and highdimensional data. To bridge this gap, the dissertation introduces a family of asymmetrical neural networks in which orthogonal transforms are embedded directly into trainable layers. These models exploit transform-domain sparsity and energy compaction while leveraging the benefits of deep network training, which adapts to the data.

Author dc:creator

• Xin Zhu (427705)

• Data compression
• Image correction
• Machine learning

Statement dc:rights

• In Copyright

DOI dc:identifier

https://doi.org/10.25417/uic.31451377.v1

OAI identifier oai:identifier

oai:figshare.com:article/31451377