Multispectral Image Denoising With a New Noise Estimation Algorithm

Multispectral images contain more spectral information about real-world scenes and are easily affected by gaussian noise when captured by sensors. Hence, denoising multispectral images is a crucial preprocessing step for various subsequent image-processing tasks, including classification, segmentation, compression, recognition, and object extraction. This article presents a novel channel-by-channel approach to denoise MSIs corrupted by Gaussian noise where each channel is subjected to a 2-level discrete wavelet transform (DWT), followed by the implementation of diverse denoising algorithms on each sub-band. However, noise does not affect each channel equally; hence, an accurate noise estimation technique is required to adaptively denoise the data. As a result, a noise estimation approach that combines DWT and singular vector decomposition is used, with the estimated variance used to determine which channels require denoising. The proposed algorithm for noise estimation and denoising is initially assessed on a LIVE dataset and then evaluated and analyzed on Sentinel-2 images. The experimental results on the multispectral data set illustrate the effectiveness of the proposed denoising technique. Experimental results on the Sentinel-2 dataset demonstrate that the proposed method outperforms the state-of-the-art methods in terms of both qualitative and quantitative analyses.