高效多尺度卷积自注意力单幅图像除雨方法

Efficient Multi-scale Convolution Self-attention Single Image Rain Removal Method

摘要:

雨天作为较常见的一种自然天气情况，会极大地影响户外视觉系统所拍摄到的图像和视频数据的成像质量并制约后续高级计算机视觉任务的性能；针对目前除雨算法存在伪影残留、细节丢失等问题，为了充分提取图像特征，有效去除雨条纹，提高除雨效率，提出一种新颖的单阶段深度学习除雨方法；采用高效卷积和跨尺度自注意力相结合的方式，弥补纯卷积网络无法满足的全局特征建模能力；嵌入多尺度空间特征融合模块，有效增加网络的感受野，增强网络对不同分布的雨条纹特征的学习能力；设计了一种混合损失函数，利用各损失函数的优势来弥补单一损失函数表现出来的缺陷；经过在不同类型数据集上的大量实验证明，该算法不仅能够有效去除雨条纹，充分保留背景细节，而且处理速度也有显著的提升。

Abstract:

As a relatively common natural weather condition, rainy days will significantly affect the imaging quality of images and video data captured by outdoor vision systems and restrict the performance of subsequent advanced computer vision tasks. In order to fully extract image features, effectively remove rain streaks, and improve the efficiency of rain removal, a novel single-stage deep learning rain removal method is proposed to solve the problems of artifacts and loss of details in current rain removal algorithms. The combination of efficient convolution and cross-scale self-attention was used to make up for the global feature modeling capabilities that pure convolutional networks cannot meet. Embedding multi-scale spatial feature fusion modules to increase the receptive field of the network effectively and enhance the network's ability to learn rain streak features of different distributions. A hybrid loss function was designed to use the advantages of each loss function to make up for the shortcomings of a single loss function. A large number of experiments on different types of datasets have proved that the algorithm can effectively remove rain streaks, fully retain background details, and has a significant increase in processing speed.