Robust destriping method based on data-driven learning

Destriping is the crucial first step of many multidetector imaging pipelines since the stripes greatly decrease the quality of obtained data and limit subsequent applications. It is also a severely ill-posed issue that estimates true gray value per pixel from a single stripe measurement. Existing approaches leverage hand-crafted filters or priors but show visually unsatisfactory results where some residual stripes still remain and the quantitative values of image data are lost. To address these problems, we propose a new data-driven method. We train a convolutional neural network on a large set of ground truth data instead of using hand-tuned filters. A UNet-like network is used to learn the regularity of complex stripe noise characteristics. To generate high-quality images, we combine a per-pixel loss and a perceptual loss to penalize mismatch between the network output and ground-truth images. Experiments show that our network significantly outperforms state-of-the-art destriping approaches in real-captured noise images of many imaging fields. Our code is available online at https://github.com/Kuangxd/DDL-SR.