融合小波变换与残差通道注意力的图像去雾算法

Image Dehazing algorithm combining wavelet transform and residual channel attention

摘要:

针对深度学习去雾算法未能充分利用非均匀分布下雾天图像的局部特征问题，提出一种融合小波变换与残差通道注意力的生成对抗去雾算法，通过构建小波-编解码器子网络和数据拟合子网络的双分支结构来复原图像。其中小波-编解码器网络利用小波变换的分解与重构代替上下采样提取图像的多尺度特征，并设计一种兼顾上下文信息的扩张卷积网络，增加了网络对纹理细节和边缘特征的捕获；数据拟合网络通过构造残差通道注意力块来增强关键特征的表达能力。此外，引入小波结构相似性损失，约束生成器输出，提高对图像内容敏感度。实验结果表明，所提算法在不同数据集上取得良好的去雾结果，且客观指标也优于大多现有算法。

Abstract:

Aiming at the problem where deep learning dehazing algorithms fail to fully utilize the local features of hazy images under non-uniform distribution, a generative adversarial dehazing algorithm that integrates wavelet transform and residual channel attention is proposed. The image is restored by constructing a dual-branch structure of wavelet encoder-decoder sub-network and data fitting sub-network. The wavelet encoder-decoder network utilizes the decomposition and reconstruction of wavelet transform instead of upsampling to extract multi-scale features of images, and an expanded convolutional network that takes into account contextual information is designed, increasing the capture of texture details and edge features; the data fitting network enhances the expression of key features by constructing residual channel attention blocks. In addition, wavelet structure similarity loss is introduced to constrain the generator output and improve sensitivity to image content. The experimental results show that the proposed algorithm achieves good dehazing results on different datasets, and the objective indicators are also found to be superior to most existing algorithms.