Self-Supervised Remote Sensing Image Dehazing Network Based on Zero-Shot Learning

Abstract

Traditional dehazing approaches that rely on prior knowledge exhibit limited efficacy when confronted with the intricacies of real-world hazy environments. While learning-based dehazing techniques necessitate large-scale datasets for effective model training, the acquisition of these datasets is time-consuming and laborious, and the resulting models may encounter a domain shift when processing real-world hazy images. To overcome the limitations of prior-based and learning-based dehazing methods, we propose a self-supervised remote sensing (RS) image-dehazing network based on zero-shot learning, where the self-supervised process avoids dense dataset requirements and the learning-based structures refine the artifacts in extracted image priors caused by complex real-world environments. The proposed method has three stages. The first stage involves pre-processing the input hazy image by utilizing a prior-based dehazing module; in this study, we employed the widely recognized dark channel prior (DCP) to obtain atmospheric light, a transmission map, and the preliminary dehazed image. In the second stage, we devised two convolutional neural networks, known as RefineNets, dedicated to enhancing the transmission map and the initial dehazed image. In the final stage, we generated a hazy image using the atmospheric light, the refined transmission map, and the refined dehazed image by following the haze imaging model. The meticulously crafted loss function encourages cycle-consistency between the regenerated hazy image and the input hazy image, thereby facilitating a self-supervised dehazing model. During the inference phase, the model undergoes training in a zero-shot manner to yield the haze-free image. These thorough experiments validate the substantial improvement of our method over the prior-based dehazing module and the zero-shot training efficiency. Furthermore, assessments conducted on both uniform and non-uniform RS hazy images demonstrate the superiority of our proposed dehazing technique.

Generative and Contrastive Combined Support Sample Synthesis Model for Few-/Zero-Shot Surface Defect Recognition

“…Further studies tried to combine knowledge graphs with zero-shot learning and achieved better performance [39]. The latest methods have also continued to apply zero-shot learning to remote sensing scene classification [40], [41], remote sensing image defogging [42], and remote sensing image super-resolution [43].…”

Remote Sensing Image Dehazing via a Local Context-Enriched Transformer

“…Attention heads in LEA are [1,2,4,8], and the number of channels are [24,48,96,192]. The decoder stage is symmetric to the encoder stage.…”

“…Remote sensing images captured by the satellite or unmanned aerial vehicle are degraded by the existing haze or cloud [1][2][3][4], which destroys the surface information acquisition and further degrades the downstream tasks including image classification [5][6][7], object detection [8][9][10], change detection [11,12], object tracking [13,14], image segmentation [15,16], and so on. Remote image dehazing methods are to recover the clean image from its haze or cloud-polluted variants, which could be applied in applications with environment monitoring, military surveillance, and so on.…”