Transfer Learning for Few-Shot Infrared Remote Sensing Image Denoising Based on Improved CycleGAN

Currently, the lack of training data gradually becomes an important factor restricting the further development of the infrared remote sensing image denoising method based on deep learning. To solve this problem, we propose a few-shot infrared remote sensing image de noising algorithm based on improved CycleGAN. The method first uses gray scale remote sensing dataset to train the network and get a pre-trained model that fully learns the features of remote sensing images. Then we adopt the Transfer learning approach to use few-shot real infrared remote sensing dataset for the second stage of training based on this pre-trained model to obtain a new model with the ability of denoising for infrared remote sensing images. In addition to this, we use a residual module containing SCConv to build the generator to reduce the number of parameters in the pre-training process, and we also add the MSE loss to the total objective function of the network to improve the retention of non-noise features in the denoising process. Experiments show that our method can effectively remove noise from infrared remote sensing images and the denoising effect is better than other similar algorithms.