Digital Image Processing and Deep Learning Approaches for Sand-Dust and Haze Image Restoration and Enhancement

Abstract

Outdoor computer vision systems utilized in conditions with severe sand dust, haze, fog, and low light face major problems that reduce image quality. These adverse conditions result in reduced visibility, color distortion, and low contrast, all of which negatively effects the accuracy and reliability of high-level vision tasks such as object detection, navigation, and surveillance. Addressing these issues is critical for enhancing the robustness of computer vision applications in real-world outdoor scenarios. This study proposes an innovative image restoration pipeline for recovering natural and visually real images affected by sand dust. To achieve superior restoration performance, the approach covers both the conventional image enhancement approaches and advanced deep learning methods using innovative transformer-based architectures. Initially, a two-stage classic image enhancement process is used to reduce color cast and enhance contrast. The first step is to remove the sand-dust-induced color cast using an intensity-corrected blue channel compensation and white balancing in the RGB color space. This method effectively recovers the natural color tone that was lost due to atmospheric scattering. The second step improves image contrast while keeping edges intact by using contrast-limited adaptive histogram equalization (CLAHE), Gaussian blurring, Laplacian filtering, and a sigmoid-based saturation modification in the HSV color space. This combination sharpens details while maintaining color balance, significantly enhancing image clarity. Building upon this, a hybrid deep learning model is introduced to further enhance restoration quality. The model includes lightweight encoders to extract local depth-wise features and Vision Transformers to capture global dependencies throughout the image. An attention fusion mechanism unifies these features, while a decoder reconstructs high-quality images with minimal inconsistencies, preserving fine details and color fidelity even under severe degradation. A key advancement in this work is the PhysFormer framework, a Physics-Guided Transformer-GAN that integrates atmospheric priors into the learning process. At its core, the Sand-Aware Transformer (SAT) employs spectral-channel attention and multi-scale adaptive fusion to model the wavelength-dependent scattering of airborne sand particles accurately. Furthermore, a novel Physics-Informed Adversarial Loss (PILoss) guides the network toward producing physically consistent and perceptually plausible results. Extensive experiments performed on both synthetic and real-world datasets, along with improved performance metrics, higher Energy Efficiency Index (EEI), and reduced time complexity, clearly shows the superior effectiveness of the proposed methods compared to existing conventional and deep learning-based methods, including SandFormer, CycleGAN, GridFormer, and MB-TaylorFormer. These improvements make the method highly suitable for deployment in resource-constrained systems such as drones, autonomous vehicles, satellites.