Image dehazing based on double branch convolution and detail enhancement

双分支卷积结合细节增强的图像去雾

School of Electronic and Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China

Received: 12 January 2024

Abstract

Because of detail loss, color distortion and contrast reduction in the image dehazing process in a haze condition, we proposed the image dehazing network based on double branch convolution and detail enhancement, which consists of image dehazing module and detail enhancement module. First, in the image dehazing module, we designed a double branch convolutional block based on depth-separable convolution and differential convolution and then combined it with the U-Net network, effectively reducing the detail loss in the image dehazing process. In the image dehazing model, we introduced the attention module composed of channel attention mechanism and pixel attention mechanism, which improves its feature extraction ability, suppresses the features that are not related to the current task and further reduces the color distortion and contrast in the image dehazing process. Then, we input the image dehazed with the image dehazing module into the detail enhancement module to further recover image details, so that the image is more similar to that in its real domain. The combination of the image dehazing module with the detail enhancement module improves the generalization ability of the image dehazing network and makes it more adaptable to the haze dataset. We carried out experiments with the public datasets ITS and Haze4K and the public real dataset IHAZE. The quantitative objective analysis and comparison show that the peak signal-to-noise ratio and the structural similarity index reach 39.69 dB and 0.994 respectively, indicating that there is a certain improvement compared with the optimal algorithm in the comparison network model. The subjective visual analysis shows that the image dehazed with the image dehazing network we proposed is more similar to the real no-haze image in terms of detail, color, contrast and so on.

摘要

针对雾天环境下图像去雾过程中出现的细节丢失、颜色失真、对比度下降的问题, 提出了一种双分支卷积结合细节增强的图像去雾网络DBDENet (double branch convolution combined with detail enhanced image de-fogging network), 此网络包含图像去雾模块和细节增强模块两部分。在图像去雾模块中, 设计了基于深度可分离卷积和差分卷积的双分支卷积块DBConv (double branch convolution), 并将其与U-Net网络相结合, 有效减轻了图像去雾过程中的细节丢失问题; 将由通道注意力机制和像素注意力机制组合成的ATT (Attenion)块引入到图像去雾模块中, 提高了模块的特征提取能力, 抑制了与当前任务不相关的特征, 进一步减轻了去雾过程中颜色失真和对比度下降问题。在细节增强模块中, 将经过图像去雾模块后的图像输入到细节增强模块中进一步恢复图像的细节信息, 使图像更趋近于真实域中的图像。图像去雾模块与细节增强模块相结合提升了网络的泛化能力, 使其在有雾数据集中具有更好的适应能力。实验在公开数据集ITS、Haze4K与公开真实数据集IHAZE上进行, 在定量客观分析比较中, 平均峰值信噪比和平均结构相似性的数值分别达到了39.69 dB和0.994, 相较于对比网络模型中的最优算法有一定提升。在主观视觉分析中, 经过DBDENet网络去雾后的图像在细节、颜色、对比度等方面相较与所提对比算法更接近于真实无雾图像。