Research on Image Inpainting Method Based on Light-Shielding Belt of Ground-Based Cloud Image

Ground-based cloud images used for ground-sky monitoring and radiation prediction suffer from key information loss due to the shadow of the lens support arm and shading strips in the images captured by the Total Sky Imager, which has unique structural features. Therefore, there is a pressing need to repair ground-based cloud images. However, existing image restoration algorithms are plagued by issues such as slow restoration speed and distorted restoration structure. To address this challenge, this paper proposes a self-supervised stepwise generation network image restoration model. The VQ-VAE model is used to compress and reconstruct the image, extract the structural and texture feature maps, and then generate the structural and texture features separately using the stepwise generation network. The generated results are subsequently fed into the decoder to obtain the final restoration result. Experiments demonstrate that the proposed image restoration method can efficiently and accurately restore cloud distribution information of the real sky, outperforming existing methods. Furthermore, it provides a promising foundation for subsequent research on meteorological changes and radiation prediction.