基于双支残差特征融合的图像去雾算法

Hazy Image Dehazing Algorithm Based on Two Branch Residual Feature Fusion

针对现有的基于卷积神经网络去雾算法无法有效地去除真实雾图非均匀分布的雾霾问题,提出一种基于双支残差特征融合网络的端到端图像去雾算法.上下文空间域注意分支针对有雾图像的高频雾气区域进行像素注意,将空间域注意模块插入多尺度扩张卷积组,对雾霾特征的像素空间进行权重赋值;通道域注意编解码分支针对高频雾霾特征的通道方向进行注意,设置ResNet自编码结构并引入通道注意解码结构对不同通道特征图的权重进行赋值;特征融合模块采用自适应权重融合像素注意和通道注意的雾层特征信息,输出不均匀雾气残差层;将原始雾图和雾气残差层作差实现图像去雾,设计判别网络提高去雾图的视觉观感.采用真实雾气图像数据集NH-Haze进行评估,实验结果表明,所提算法对非均匀分布雾图的去雾视觉效果良好,在峰值信噪比和结构相似度评价上均优于对比算法.

Aiming at the problem that the existing dehazing algorithms based on convolution neural network can not effectively remove the non-uniform distribution of haze in the real-haze image,an end-to-end image dehazing algorithm based on two branch residual feature fusion network is designed.The attention branch in context spatial domain pays attention to the high-frequency hazy region of hazy image.The spatial attention module is inserted into the multi-scale expansion convolution group to assign the weight to the pixel of hazy characteristics.The channel attention branch pays attention to the channel of high-frequency hazy features,sets ResNet self-coding structure,and introduces channel attention decoding structure to assign the weight of different channel feature maps.The feature fusion module uses adaptive weights to fuse the hazy layer fea-ture information of pixel attention and channel attention,and outputs the uneven hazy residual layer.The original hazy image and the hazy residual layer are compared to achieve image dehazing,a discrimination network is designed to improve the visual appearance of the dehazing image.The experimental results of evaluating the algorithm with real-haze images NH-Haze show that the algorithm has good visual effect on non-uniform hazy images,and is superior to other comparison algorithms in peak signal-to-noise ratio and structural similarity evaluation.