【摘要】 煤矿井下工作环境恶劣，受复杂光照及尘雾等影响监控图像模糊不清，很大程度上影响煤矿作业的安全性。现有合成尘雾数据集方法假定大气光是均匀分布的前提条件在煤矿场景中并不成立，使得深度去雾算法在煤矿场景中泛化性能较差。针对该问题，首先估计清晰图像深度信息，利用Retinex理论得到对应带雾图像空间变化的亮度信息，结合图像深度信息和亮度信息通过神经网络训练，在亮度一致性等损失约束下生成能够反映空间亮度变化的带雾图像。另外，考虑到煤矿真实场景光照复杂的特点，对清晰图像进一步增强处理，消除图像去雾后过于昏暗的问题。结合煤矿真实场景的对比实验，表明了本文方法的有效性。更多还原

【Abstract】 The working environment underground in coal mines is harsh, and the monitoring images are blurred due to complex lighting and dust mist, which greatly affects the safety of coal mining operations.The existing methods for synthesizing dust and mist datasets assume that the atmospheric light is uniform, which does not hold in coal mining scenarios, resulting in poor generalization performance of deep defogging algorithms in coal mining scenarios.In response to this issue, the depth information of clear images were first estimated, and used Retinex theory to obtain the spatial brightness information of fogged images.Combining the depth information and brightness information of the images, a neural network was trained to generate hazy images that can reflect the spatial brightness changes under loss constraints such as brightness consistency.In addition, considering the complex lighting in the real scene of coal mines, clear images were further enhanced to the dim artifacts after image dehazing.Experimental results demonstrate that the proposed method is suitable for synthesizing the realistic haze/clear dataset in coal mine.更多还原