LGASR: latent-content guided adversarial sand-dust image reconstruction strategy

The research on sand-dust image enhancement usually follows the developmental dynamics of image haze removal and is transitioning from traditional methods to end-to-end (e2e) learning-based algorithms. However, the more complex degradation of sandstorm images inevitably increases the potential risks in e2e algorithms, leading to unstable model performance. To bridge this gap, we reanalyze the extractor-reconstructor structure and propose a latent-content guided adversarial sand-dust image reconstruction (LGASR) strategy. Specifically, LGASR alternates the training of the backbone network between the learning line and the guiding line, optimizing the extractor to accurately capture the latent content of input images and enabling the reconstructor to reconstruct target images based on the extracted content. Additionally, we designed a module named Desandformer to enhance the model’s ability to extract and utilize latent features. Experimental results on both synthetic and real-world sandstorm images demonstrate the superior performance of LGASR.