基于非局部信息的大气偏振模式生成方法

A generative network based on non-local information for atmospheric polarization modelling

大气偏振模式作为一种稳定的自然属性,因其包含具有方向信息的∞字形特征和子午线特征,在导航、探测等领域具有广泛地应用。针对实际天气条件下,大气偏振信息获取方式受限于动态云层的干扰,导致大气偏振模式的分布规律遭到破坏和部分信息缺失的问题,提出一种基于非局部信息的大气偏振模式生成方法,并设计了一种非局部信息修复模块进行两个阶段的修复,第一阶段通过挖掘不同区域的大气偏振模式空间分布的关联性和全局性,实现对缺失区域空间维度上偏振信息的修复;第二阶段利用大气偏振信息在不同时刻的特征映射关系和分布连续性,实现对云层干扰区域时间维度上特征信息的修复。在Temporal Polarization 1072偏振数据集上的实验结果定性和定量的表明,该方法能够有效地去除大气偏振模式中的云层干扰噪声,修复缺失区域的偏振信息,同时生成的结果具有更高的结构一致性和语义一致性。

As a stable natural attribute,the atmospheric polarization mode is widely used in various fields such as navigation and detection because it contains the∞shape feature and meridian feature with directional information.However,obtaining atmospheric polarization information in real weather conditions is a challenging task due to the limitations imposed by dynamic cloud interference.This limitation causes the distribution law of atmospheric polarization information to be destroyed and,in turn,leads to the loss of some information.To address this issue,we proposed an atmospheric polarization mode generation method based on non-local information.Then,a non-local information inpainting block was designed for two-stage repair.In the first stage,the non-local spatial continuity information of the atmospheric polarization mode was mined to enhance the global structure between feature information and realize spatial information repair.In the second stage,the feature mapping relationship was established between atmospheric polarization information at different times,and the time continuity of the non-local atmospheric polarization information distribution was employed to repair feature information of superimposed noise regions in the time dimension.The experimental results on the Temporal Polarization 1072 polarization dataset qualitatively and quantitatively demonstrated the efficacy of this method in effectively removing cloud interference noise in the atmospheric polarization mode and repairing polarization information of missing areas,and higher structural and semantic consistency of the generated results.