基于通道先验损失的无监督图像去雾算法

Unsupervised Image Dehazing Algorithm Based on Channel Prior Loss

图像去雾是图像处理的一个关键步骤.基于学习的方法由于在收集清晰和模糊图像的固有限制,通常依赖于合成数据进行训练,且由室内图像和相应的深度信息构成.在处理室外场景时,可能会存在域偏移问题.提出了一种完全无监督的训练方法,通过最小化暗通道先验能量函数来进行图像去雾.此外,只使用真实世界的室外图像进行训练,并通过直接最小化该能量函数来优化网络参数.实验结果表明该方法的性能与大规模监督方法相当,可见通过网络和学习过程,可以获得额外的正则化.

Image dehazing is a key step in image processing. Due to the inherent limitation of collecting clear and blurred image pairs, learning-based methods usually rely on synthetic data for training, and are composed of indoor images and corresponding depth information. In dealing with outdoor scenes, domain offset problems may occur. A completely unsupervised training method was proposed, which minimized the dark channel prior energy function for image dehazing. In addition, only real-world outdoor images were used for training, and the network parameters were optimized by directly minimizing the energy function. The experimental results showed that the performance of the proposed method was comparable to that of the large-scale supervision method. In conclusion, additional regularization could be obtained through the network and learning process.