改进多尺度卷积神经网络的单幅图像去雾方法

针对当前已有的去雾方法容易造成天空区域存在光晕以及色彩失真的现象,提出了一种多尺度卷积结合大气散射模型的单幅图像去雾算法。将原始有雾图像与三个不同尺度的卷积核进行卷积,经过一系列特征学习后得到粗略的传播图,然后使用引导滤波器对其进行优化,得到精细化后的传播图。利用粗传播图和有雾图像计算出全局大气光。根据大气散射模型反推出无雾清晰图像。实验结果表明,该方法对天空区域的处理更加自然,在图像的纹理细节以及颜色失真上有较好的效果。

Single Image Dehazing by Using Improved Multi-Scale Convolutional Neural Network

As the current reported dehazing method is easy to cause halo and color distortion in the sky region, a single image dehazing algorithm by combining multi-scale convolution with scattering model is proposed. Firstly, the original haze image is convoluted with three different scales of convolution kernels. After a series of characteristic learning, the rough transmission is obtained. Then the transmission map is refined by using the guided filter. Secondly, according to the haze image and rough transmission, the global atmospheric light is known. Finally, with the refined transmission map and the calculated atmospheric light, the final dehazed image is inversely derived from the atmospheric scattering model. Experimental results show that the proposed algorithm is more natural to deal with the sky area, and it has better restoration effect on image texture and color distortion.