A Spatial-Frequency Based Multiscale Destriping Network for Remote Sensing Images
Feiyan Wu, Bo Fu, Yu Shi · 2024
Stripe noise in remote sensing images significantly reduces image quality. Therefore, effectively removing stripe noise is crucial for enhancing the application value of remote sensing images. The striping noise of remote sensing images usually has certain directional characteristics, which can be represented in the frequency domain through discrete wavelet transform (DWT). At present, few networks consider simultaneously fusing spatial and frequency features for learning, thus there is still room for improvement in the ability to remove stripe noise. To improve this, we propose a spatial-frequency based multiscale destriping network for remote sensing images (SFBMDNet). We construct a frequency-spatial convolution block (FSCB) as the building block of the encoders and decoders, which introduces DWT to obtain frequency feature and uses convolutions to capture spatial feature, and then combines the two types of features. Additionally, to fully utilize the features learned by the encoders, we develop a spatial-channel attention fusion block (SCAFB). Our method is compared with state-of-the-art methods on simulated and real remote sensing images, and the results show that the proposed SFBMDNet has good performance in removing image stripes, and outperforms the comparative methods in terms of visual quality and metrics.