基于多尺度卷积神经网络的图像去雾方法研究

Research on Image Defogging Method Based on Multi-Scale Convolutional Neural Network

摘要: 针对图像去雾算法中存在光晕、失真或去雾不彻底等问题，本文提出了一种基于多尺度卷积神经网络的去雾方法。首先，利用多个较小的单尺度卷积提取浅层信息，其次通过多尺度卷积并行提取得到多维度的深层特征信息，增加了网络深度，丰富了语义信息，将浅层特征和深层特征进行特征融合，最后利用非线性回归来生成透射率特征。在RESIDE数据集中的ITS和OTS子数据集进行了实验分析，本文算法的SSIM和PSNR分别为0.976、25.041，结果表明，本文算法图像恢复度更佳，对易失真区域及纹理细节的处理更加自然，在去雾效果上较传统算法更好，且该方法在主观及客观上均有更好的评价。

Abstract: Aiming at the problems of distortion, halo, or incomplete defogging in image defogging algorithms, this paper proposes an image defogging method based on a multi-scale convolutional neural net-work. First, several smaller single-scale convolutions are used to extract shallow information, and then multi-scale convolutions are used to extract multi-dimensional deep feature information in parallel, which increases the depth of the network, enriches semantic information, fuses shallow features and deep features, and finally uses nonlinear regression to generate transmittance fea-tures. Experimental analysis was conducted on the ITS and OTS sub-datasets in the RESIDE dataset. The SSIM and PSNR of the algorithm proposed in this paper were 0.976 and 25.041, respectively. The results showed that the algorithm proposed in this paper had better image restoration, more natural processing of easily distorted areas and texture details, better defogging performance than traditional algorithms, and better subjective and objective evaluation.