Deep learning-based algorithm for synthetic aperture radar despeckling

• Main Author: Wang, Yuxuan
• Other Authors: Teh Kah Chan
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Telecommunication engineering
• Online Access: https://hdl.handle.net/10356/181410
• Institution: Nanyang Technological University

Synthetic Aperture Radar (SAR) possesses the unique advantages of being all-weather and all-day operational, which optical images cannot substitute, making it highly valuable in both military and civilian applications. However, SAR images inherently suffer from speckle noise, which severely hinders the effective implementation of subsequent tasks such as image segmentation and target detection. Therefore, research on SAR image denoising and target detection technologies is of great significance. In recent years, the rapid development of deep learning technologies has produced numerous outstanding models that have achieved remarkable success in natural image processing. These advancements provide new insights for SAR image denoising and target detection. This dissertation focuses on deep learning-based SAR image denoising and target detection techniques, with the following specific contributions and innovations. This dissertation provides a detailed introduction to the imaging mechanisms, models, and current research status of SAR image denoising. It also analyzes the advantages, disadvantages, and denoising performance of the current mainstream algorithms. A deep learning-based SAR image denoising network, DespeckleNet, a kind of convolution neural network was constructed, and its denoising performance was tested using both simulated and real data. The results were compared with several classical algorithms. Additionally, this dissertation proposes a SAR image speckle noise suppression model based on a residual optimization network. This model utilizes a residual optimization strategy to construct a learning model for the speckle noise residual image, achieving an optimal identity mapping for speckle noise suppression. Experimental results on simulated datasets, real datasets, and target detection methods demonstrate that the proposed model not only effectively suppresses speckle noise but also preserves the local edge details of SAR images. Significant improvements were achieved in both detail preservation and noise suppression, laying a solid foundation for subsequent SAR target detection, recognition, and typical target monitoring. Keywords: Synthetic Aperture Radar, Convolution Neural Network, Denoising Network, DespeckleNet, Residual Learning.