Implementing Wavelet Transforms in the Pooling Process of Convolutional Neural Networks for Image Denoising

Neural networks (NNs) have been shown to bring powerful image denoising performance. The demand for performance improvement brings the desire for a larger receptive field size. Obtaining a larger receptive field size by pooling becomes an option. However, the typical pooling method process has information loss, which is unacceptable for image noise reduction. Previous studies have tried to introduce wavelet methods to expand the receptive field with no loss of information. So far, only Haar wavelets have been successfully used for this purpose. In this work, other wavelet transforms, and inverse wavelet transforms are implemented, and their performance is evaluated on autoencoder and Multi-level wavelet convolutional neural networks. The results show that wavelet methods have a strong potential for image denoising, while the network structure that can fully exploit wavelets' performance is yet to be improved.

Author

• Fu, Weimin

Language

• en

Date created

• 2021

Type of Work

• Master's Thesis

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Electrical Engineering

Degree

• M.S.

Advisor

• Doroslovacki, Milos

Committee Member(s)

• Mahdi, Imani
• Ahmadi, Shahrokh

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/xd07gt43w