|
چکیده
|
The lack of paired training data poses a signifi cant challenge for the data-driven sand-dust image enhancement models. Supervised methods rely on simulated data, but the domain gap between simulated and real-world scenarios limits their generalizability. Unsupervised methods, while avoiding this issue, often have complex ar- chitectures and fail to restore fi ne details. Motivated by these, we propose a semi-supervised sand-dust image enhancement method, SSDIE-Net, which integrates the strengths of supervised and unsupervised learning within a unifi ed framework. SSDIE-Net is trained on simulated data using supervised reconstruction loss functions in the supervised branch. It integrates classical image restoration techniques with conditional adversarial networks to generate highly realistic sand-dust images. Moreover, SSDIE-Net adopts consistency regularization, dark chan- nel priors-based regression minimization, Retinex-based pseudo-labeling, and adversarial learning to translate sand-dust images to clean ones in the unsupervised branch. Additionally, we designed an attention-based multi- scale feature fusion network in which feature map extraction from diff erent scales facilitates improved local-to- global learning. Unlike previous methods that focus on extracting local features, SSDIE-Net learns long-range dependencies, which are essential for understanding the overall scene structure. Extensive experiments show that SSDIE-Net outperforms state-of-the-art supervised, unsupervised, and classical methods, producing dust- free images with enhanced details and better generalization to real-world scenarios. The code is available at https://github.com/bartani/SSDIE-Net
|