基于在线知识蒸馏的物理感知图像去雾算法

Physically-aware Image Dehazing Algorithm Based on Online Knowledge Distillation

为解决传统图像模型在去雾时会产生伪影、光晕等问题,结合深度学习,提出了一种基于在线知识蒸馏的物理感知图像去雾算法。算法首先通过物理感知单元构建的多尺度网络获取丰富的共享特征,然后采用基于模型的方式和端到端的方式分别生成去雾图像,最后采用特征聚合块来融合两种方式生成的去雾图像。此外,算法采用在线知识蒸馏的学习策略来联合优化网络。实验结果表明:与其他去雾算法相比,所提算法在合成图像和真实场景图像上均取得了优异性能,在合成有雾室外数据集SOTS上的峰值信噪比和结构相似度分别为23.59 dB和0.937。

To solve problems of artifacts and halos generated by traditional image models in dehazing,a physically aware image dehazing algorithm was proposed based on online knowledge distillation and deep learning.First,shared features were acquired by a multi-scale network constructed by physically-aware units.Then,dehazed images were generated by model-based and end-to-end approaches.Finally,feature aggregation blocks were employed to fuse the dehazed im-ages generated by the two approaches.Additionally,the learning strategy of online knowledge distillation was introduced to jointly optimize the network.Results showed that the proposed algo-rithm achieves excellent performance on both synthetic and real scene images than other dehazing algorithms.On the SOTS outdoor dataset,it achieves a PSNR of 23.59 dB and SSIM of 0.937.