Transformer-Based Feature-Preserving Despeckling Framework for Synthetic Aperture Radar Imagery

Synthetic Aperture Radar (SAR) images are regularly used in the remote sensing because these images can be generated in any weather or light condition. However, the SAR images are always exaggerated by speckle noise, which makes the SAR images less clear and makes it difficult to get useful structural and textural data for processing. To make it effective, despeckling methods must minimize noise while retaining the important image features such as edges and fine textures. To solve these issues, this article developed a Transformer Based Despeckling Network (TBDN). The developed framework initially extracts low level features from the noisy SAR image by use of convolutional layers. Then, transformer encoder blocks use extracted features with the help of multi head self-attention mechanisms to achieve long range spatial dependencies as well as contextual data. Hierarchical characteristics are collected in a multi-scale feature fusion module to support preserve the structure while reducing noise. Finally, the despeckled SAR image is enhanced with the help of image reconstruction layer.