基于改进Ghost模块的快速去雾算法

Fast defogging algorithm based on improved Ghost module

最近基于卷积神经网络的端到端学习算法在去雾方面取得了进展,但这些算法大都使用过深过大的网络去拟合含雾图像数据,从而导致去雾速度偏慢.为解决此问题,本文提出了一个用于单幅图像快速去雾的高效去雾模型.该模型基于改进的Ghost模块构建轻量级神经网络;将金字塔池化和注意力机制结合获取图像不同区域的上下文信息,从而提高了网络获取含雾图像全局信息的能力.该网络在RESIDE的室外数据集进行训练和测试.实验结果表明:相比较DCP、MSCNN、AODNet等先进算法,所提出的方法具有令人满意的去雾质量和速度.

Recently,end-to-end learning algorithms based on convolutional neural networks have made progress in defogging,but most of them use networks that are too deep and too large to fit foggy image data,resulting in slow defogging speed.To solve this problem,an efficient defogging model is proposed in this paper,which can be used to quickly defog a single image.The model is based on the improved GHOST module to build a lightweight neural network.Pyramid pooling and attention mechanism are combined to obtain the context information of different areas of the image,thus improving the network's ability to obtain the global information of foggy images.The network RESIDE in Donatella's outdoor data set for training and testing.Experimental results show that,compared with DCP,MSCNN,AODNet and other advanced algorithms,the proposed method has satisfactory defogging quality and speed.