A Denoising Method for Salt and Pepper Noise in Remote Sensing Based on Swin-Transformer Convolution U-Net and Filtering—FSCU-Net

Remote sensing imagery has become an indispensable tool for cost-effectively capturing extensive geospatial data across diverse applications. However, this technology remains fundamentally susceptible to noise contamination. Salt and pepper noise is one of the common issues that can significantly impair image quality and hinder subsequent processing tasks. While numerous methods have been proposed to mitigate this noise, many traditional techniques result in the loss of critical image detail. Recent advances in deep learning-based denoising approaches have shown considerable promise in addressing this challenge. A notable framework is based on using the Swin-Transformer Convolution U-Net, which effectively integrates Swin-Transformer and convolutional layers to enhance salt and pepper noise removal while minimizing signal loss. However, denoising performance may decline under high noise density conditions, leading to visible color discrepancies. To overcome this limitation, we introduce FSCU-Net, a hybrid approach that combines traditional denoising techniques with deep learning methods. FSCU-Net employs cyclic switching mean filtering for initial noise reduction, followed by a Swin-Transformer Convolution U-Net for further processing. As a novel architecture in computational imaging, FSCU-Net establishes the integration framework combining cyclic switching mean filters with Swin-Transformer modules. This approach embodies a paradigm shift from conventional physics-based denoising methodologies by its exclusive reliance on data-driven learning mechanisms unconstrained by prior physical models. Notably, the adopted filtering operations serve solely as preliminary quality enhancement components prior to feature extraction. Experimental results demonstrate that FSCU-Net significantly improves the denoising performance and reduces the color discrepancies in the presence of high-density salt and pepper noise.