Abstract
Denoising remote sensing images is important for their subsequent use and research. To meet the demands of image denoising in strong noise environments, this study proposes an improved BM3D remote sensing image denoising network (CBMNet) based on a context aggregation network. Initially, the original image is read, and the image outputted by BM3D filtering is stored. Following this, random patches extracted from the original image are fed into the network alongside the image processed through BM3D filtering to extract data, which is then input into the network. Subsequently, multi-scale CAN layers are established to calculate the l_2 loss function between the standard output of the BM3D filter and the network response post-processing the input image using the CAN network. Ultimately, the CBMNet network is trained to approximate the BM3D filtering operator. The experimental results indicate that, both in terms of subjective visual assessment and objective evaluation metrics, the proposed method outperforms the classical BM3D algorithm as well as the Wiener, mean, and Gaussian filtering denoising methods in removing strong Gaussian noise from remote sensing images. Additionally, the proposed method better preserves image edge details and texture information, resulting in clearer image outputs. This has significant reference value and practical utility for subsequent application research in remote sensing imagery. These results bear substantial reference significance and practical utility for subsequent applications in remote sensing imagery research.