基于多尺度信息交互的深度学习图像去沙算法

Depth learning image desanding algorithm based on multi-scale information interaction

沙尘天气下,入射光线的散射和吸收,会导致图像退化,出现色彩偏移、细节丢失等问题,户外计算机视觉系统的工作性能受到严重影响。为此,提出了一个端到端的基于多尺度信息交互(multi-scale information interaction,MSII)的网络结构。该网络采用并行的两个不同分辨率子网,通过上下采样使两个子网信息交互,引入交叉注意力机制进行空间、特征融合,以获得更丰富的细节;提出了一个简单有效的沙尘合成方法,并以此构建了一个配对沙尘数据集。实验可得,与所比较方法中最好的结果相比,在合成数据上,结构相似度提高5.94%,峰值信噪比提高0.403 dB;在真实数据上,自然图像质量指标提高0.4407,对比度、标准差、信息熵分别提高1.5315、1.0152、0.3352。由此可知,所提方法可获得细节清晰且色彩鲜明的图像。

In sandstorm weather,the scattering and absorption of incident light can lead to image degradation,color deviation,and loss of details,which seriously affects the performance of outdoor computer vision systems.To this end,an end-to-end network structure based on multi-scale information interaction(MSII)is proposed.This network adopts two parallel subnets of different resolutions,which interact with each other through upsampling and downsampling,and introduces a cross attention mechanism for spatial and feature fusion to obtain richer details.A simple and effective method for sand and dust synthesis was proposed,and a paired sand and dust dataset was constructed based on this method.The experiment shows that compared with the best results among the compared methods,the structural similarity is increased by 5.94%and the peak signal-to-noise ratio is increased by 0.403 dB on the synthesized data;In real data,the natural image quality index is improved by O.4407,and the contrast,standard deviation and information entropy is increased by 1.5315,1.0152 and 0.3352.From this,it can be seen that the method proposed in this article can obtain images with clear details and bright colors.