Visibility enhancement of sand-dust images based on depthwise separable convolution

To address the visual interference problem in 'sand-blind' environments, where sand images suffer from color shifts and severe loss of background detail, we propose a sand image enhancement algorithm based on color correction and depthwise separable convolution. First, the color correction module preprocesses the synthetic sand images to correct color deviations. Then, the SMU (Separable Multi-scale Unit) is applied within the decoupling reconstruction module to address gradient vanishing and extract degradation information. Finally, aggregated depthwise separable convolutions are added to the denoising module to learn textures, thereby obtaining more effective information and target features. Experiments on synthetic datasets demonstrate that, compared to mainstream algorithms such as TFIO, GDCP, AWC, and HRDCP, our algorithm achieves at least 10.46% and 35.56% improvements in PSNR and SSIM metrics, respectively, better restoring image details obscured by sand.