基于倍频卷积和注意力机制的图像去雨

针对机器视觉场景图像中由于雨线影响导致背景信息模糊、损失的问题,提出一种基于倍频卷积和注意力机制的图像去雨方法.首先,建立基于空-频域去雨模型,设计基于空间尺度变换和倍频卷积的频率特征分解模块,通过学习得到频率特征和雨线特征的映射关系,降低低频特征空间冗余,提高网络运行效率;其次,设计多层通道注意力模块映射雨线层权重信息,增强重要特征,挖掘雨线层之间的亮度差异,提高雨线检测性能;最后,通过序列操作迭代分解出不同成分的雨线信息,进而完成场景图像去雨.实验结果表明,所提方法对不同方向、形状的雨线和雨滴具有良好的去除性能,同时对于背景图像的细节与边缘信息也具有较好的保护作用.

Image de-raining based on octave convolution and attention mechanism

Aiming at the problem of background information blur and loss caused by rain streaks in machine vision scene images, a method of image de-raining based on octave convolution and attention mechanism is proposed. Firstly, a de-raining model based on spatial-frequency domain is established, and a frequency feature decomposition module is designed based on spatial scale transformation and octave convolution. The mapping relationship between frequency features and rain streaks features is obtained through learning, so as to reduce spatial redundancy of low-frequency features and improve network operation efficiency. Secondly, the multi-layer channel attention module is designed to map the weight information of rain streaks, enhance important features, mine the brightness difference between rain streaks layers, and improve the performance of rain streaks detection. Finally, the rain streaks information of different components is decomposed iteratively through sequence operation, and then the de-raining of scene image is completed. Experimental results show that the proposed method has good removal performance for rain streaks and raindrops with different directions and shapes, and also has good protection for details and edge information of background image.