Separable Multi-scale Large Kernel Convolutional Remote Sensing Denoising Network

Abstract

Abstract: The abstract of the study stated that remote sensing images contain abundant details of land objects and terrain, and the denoising process should strive to preserve these critical pieces of information. However, traditional CNN methods performed poorly when dealing with high-resolution, multi-scale, and complex scenes, as they struggled to model the long-range dependencies within images. Methods based on Transformer improved this issue through the self-attention mechanism; however, their high computational cost limited their application in resource-constrained environments. To address this, a Multi-Scale Large Kernel Detail Enhancement Network was proposed, aiming to effectively retain the detailed information in remote sensing images. By utilizing pooling to separate high and low-frequency information, the approach adopted separable multi-scale large kernel convolutions to capture extensive spatial information, enhancing high-frequency features while reducing computational complexity. These innovative techniques effectively expanded the receptive field, improving the denoising effect of remote sensing images. Currently, compared with the best results from other methods, MLKNet achieves an average improvement of approximately 3.1 dB in grayscale remote sensing image denoising across three different noise levels, and an average improvement of about 1.17 dB in color remote sensing image denoising under the same conditions.