Unsupervised Denoising with Implicit Noise Mapping for Single Martian Multispectral Image
Junjie Li, Weikun Lv, Jiawei Wang, Yumei Wang, Yü Liu · 2024
Multispectral (MS) images of Mars are extensively used for material recognition and classification. However, due to constraints like limited lighting, photon issues, and atmospheric interference, these images inevitably suffer from noise, which greatly hinders their further applications. The scarcity of MS images and limited research on their inherent noise characteristics make denoising a more challenging task. In this paper, we employ a novel unsupervised estimation technique to quantify noise, which does not depend on prior knowledge of distributions, thereby enabling implicit modeling of the noise. By integrating the quantified noise characteristics into denosing model with attention mechanism, the model’s ability to perceive and adapt to noise is significantly enhanced. Experiment shows that our method not only proves effective in denoising raw Mars MS images but also demonstrates competitive performance compared to state-of-the-art methods.