融合压缩激活注意机制的图像去雾算法

Research on the image dehazing algorithm of the Squeeze-and-excitation mechanism

针对非均匀带雾图像出现颜色失真和细节丢失的问题,提出一种融合压缩激活注意力机制的端到端感知去雾卷积神经网络。首先,根据非均匀雾图特征,在特征融合注意网络上引入压缩激活注意力机制,通过学习的方式自动获取每个特征通道的重要程度,并对其进行排序后去提升有用的特征权重,遏制对当前任务用途较小的特征比例;其次,在损失函数方面融入感知损失,使去雾网络模型学习到更多语义特征信息,从而加强被模糊的边缘细节,获取较好的去雾效果;最后,在非均匀带雾图像NH-HAZE数据集上进行定性和定量分析。实验结果表明,所提出的网络与经典方法相比在量化指标PSNR、SSIM分别提升了3.05 dB和0.08%,且主观视觉效果上保留了更多的边缘信息和纹理细节。

In response to the problem of color distortion and loss of details in non-uniform haze images,this article proposes an end-to-end perception of the haze sense of fusion and activation mechanism.First of all,according to the characteristics of non-uniform haze maps,introduce compression-activation mechanisms on the network,automatically obtain the importance of each characteristic channel through learning,and sort it to improve the useful feature weight of useful features to curb The characteristic ratio of the current task is small;secondly,integrate the loss of perception in the aspect of the loss function,so that the dehazing network model learns more semantic special information,thereby strengthening the blurred edge details and obtaining better haze effects;in the end,Bling and quantitative analysis on the non-uniform haze image NH-HAZE dataset.The experimental results show that the network proposed in this article has increased 3.05 dB and 0.08%compared with the classic method on PSNR and SSIM,respectively,and more edge information and texture details are retained on the subjective visual effects.