基于多补丁和多尺度层级聚合网络的快速非均匀图像去雾

Multi-patch and Multi-scale Hierarchical Aggregation Network for Fast Nonhomogeneous Image Dehazing

尽管基于卷积神经网络的去雾算法在合成的均匀雾气数据集上已经取得了巨大进展,但在真实的非均匀有雾图像上仍然表现不佳。为了快速有效地去除图像中的非均匀雾气,文中首先提出了一种多补丁和多尺度层级聚合网络结构(Multi-patch and Multi-scale Hierarchical Aggregation Network,MPSHAN),融合了多补丁局部化信息和多尺度全局化信息。其次,提出了层级融合模块(Hierarchical Fusion Module,HFM),既解耦了残差融合以实现更丰富的非线性特征表达,又通过通道注意力机制提升了关键位置的特征融合质量。同时,对层级结构使用扩张卷积获得多尺度信息,增强特征图以优化融合效果。此外,在损失函数中加入频域损失以恢复更好的边缘质量。实验结果表明,所提算法在非均匀雾气图像上具有很好的鲁棒性,1200×1600高分辨率图像的平均处理时间仅有0.044 s,相比其他去雾算法,其在图像去雾效果和运行时间之间实现了更好的平衡。

Despite dehazing algorithms based on convolutional neural networks have made tremendous progress in synthetic uniform hazy datasets,they still perform poorly on real nonhomogeneous hazy images.In order to achieve fast and effective nonhomogeneous image dehazing,we propose a multi-patch and multi-scale hierarchical aggregation network(MPSHAN),which fuses multi-patch local information and multi-scale global information.Secondly,we propose a hierarchical fusion module(HFM),which not only decouples residual fusion to achieve richer non-linear feature expression,but also improves the feature fusion quality at key locations through the channel attention mechanism.At the same time,dilated convolution is used on hierarchies to obtain multi-scale information,which enhances feature maps to optimize the fusion effect.In addition,in the loss function,we add frequency domain loss to restore better edge quality.The experimental results show that the proposed algorithm has good robustness on nonhomogeneous hazy images,and the average processing time of 1200×1600 high-resolution images is only 0.044 s.Compared with other dehazing algorithms,it achieves a better balance between image dehazing effect and running time.