摘 要：近年来，基于深度神经网络的航空图像去雾技术取得了显著进展。然而，大多数方法仅学习从雾霾图像到清晰图像的单向映射，未充分利用雾霾生成过程中的潜在监督信息，导致去雾性能下降。此外，现有在特征空间引入大气散射模型的方法忽略了传输图与大气光之间的物理负相关性，这对学习更具代表性的特征至关重要。为此，本文提出了一种双向映射与物理负相关的航空图像去雾算法。具体来说，本文构建了双向映射多阶段框架，在共享参数的网络中协同建模雾霾去除与生成过程，有效探索了雾霾形成机制，从而提升去雾性能。此外，设计了大气光引导的物理负相关感知模块，通过负相关引导机制利用大气光信息优化传输图特征估计。在公开航空雾霾数据集上的实验结果表明，本文方法在客观评价指标和视觉效果方面均优于当前最先进的7种去雾算法。

Abstract: Recently, deep neural networks based aerial image dehazing techniques have been achieved remarkable per-formance. However, most methods only learn a one-way mapping from hazy images to clear images, which fail to fully exploit the latent supervision information of haze generation, leading to dehazing performance degradation. Moreover, ex-isting methods that incorporate the atmospheric scattering model in the feature space overlook the exploration of the nega-tive correlation between the transmission map and atmospheric light, which is crucial for leaning more representative fea-tures. To address these issues, we propose a Bidirectional Mapping Physics-Aware Negative Correlation algorithm for aerial image dehazing. Specifically, the proposed Bidirectional Mapping Multi-Stage Framework synchronously models haze re-moval and generation processes in a parameter-sharing network, effectively exploring the haze formation mechanism to enhance dehazing performance. Furthermore, an Atmospheric Light-Guided Physics-Aware Negative Correlation Module is designed, which leverages a negative correlation guidance mechanism to optimize the estimation of transmission map features utilizing atmospheric light information. Experimental results on public datasets demonstrate the proposed dehaz-ing algorithm outperforms seven state-of-the art haze removal methods in both objective evaluation and visual quality.