Remote Sensing Image Denoising Based on Multi-Scale Feature Fusion and Regional Contextual Information

Abstract

The various types of noise in Remote Sensing (RS) images resulting from environmental factors and the imaging system, often significantly degrade the imaging quality and impair high-level visual tasks. Therefore, denoising plays an essential role in the applications of RS images. Traditional methods mainly focus on dealing with a single type of noise, while the denoising performance is rather limited with the complex noise in practice. Given the advanced representation learning ability of deep neural networks, investigations have been made to apply them to the RS image denoising. However, existing methods tend to pay more attention to global features, the detailed local information are often overlooked. Therefore, in this paper, we propose a hyperspectral RS image denoising model by leveraging both multi-scale feature fusion and regional contextual information. More specifically, the proposed model includes two branches, i.e., the global branch based on Multi-scale Feature Fusion Module (MFFM) to aggregate global features from multiple scales and the local branch based on Transformer Attention Module (TAM) to explore the regional context. The denoised RS image is then obtained by combining feature maps from two branches. Extensive experimental results demonstrate that our model performs favorably on both simulated and real noisy RS images. The proposed model is also evaluated on the high-level visual tasks including object detection and clustering, which further illustrates the potential of our model for facilitating downstream tasks.