SAR Image Generation Using a Modified CycleGAN

Abstract

The important issue of synthetic aperture radar (SAR) image is the amount of dataset, because it is hard to obtain the image suitable for the purpose. This shortcoming can be complemented through using deep learning. Deep learning processes a large amount of data and performs the desired research, which is used in various fields. This thesis propose the method using deep learning model to generate images similar to original optical and SAR images, and we adopt CycleGAN as our model. CycleGAN is a method to generate images by using unpaired datasets. To remove speckle noise, we apply several filters to enhance the accuracy of the model. Moreover, this thesis uses other deep learning models to compare the precision, and performs quantitative analysis using several indexes. Experiment result shows that the generated images maintain the features of original images, and filtered images are more similar to original SAR images than non-filtered images.