基于深度频率特征注意力机制的图像去雨方法  

Image de-rain based on deep frequency feature attention mechanism

作　　者：杨青 于明 阎刚 YANG Qing;YU Ming;YAN Gang(College of Electronic Information Engineering,Hebei University of Technology,Tianjin 300401,China;Shijiazhuang Campus of Army Engineering University,Shijiazhuang 050003,China;School of Artificial Intelligence,Hebei Universityof Technology,Tianjin300401,China)

机构地区：[1]河北工业大学电子信息工程学院,天津300401 [2]陆军工程大学石家庄校区,石家庄050003 [3]河北工业大学人工智能与数据科学学院,天津300401

出　　处：《吉林大学学报（工学版）》2024年第10期2952-2962,共11页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(61806071,62102129)。

摘　　要：针对雨图像中由于雨线和雨幕效应导致背景信息模糊、清晰度下降的问题,本文提出一种基于深度频率特征注意力机制的图像去雨方法。首先,构建基于频率去雨模型的参数估计网络,设计基于空间尺度变换和倍频卷积的频率特征分解模块区分参数频率特征,降低低频特征空间冗余,提高运算效率。其次,在雨线检测模块中利用多频通道注意力机制映射雨线层权重信息,增强权重特征多样性,提高雨线检测性能。最后,设计基于焦频损失的去雨修复网络,进一步修复去雨模型估计图像,通过缩小频率差距提高图像修复质量,增强网络的抗干扰能力。实验结果表明,所提方法能够有效去除真实雨图像场景中的雨线,抑制雨幕效应,保留图像背景信息完整,图像清晰度较高。Aiming at the problem of blurred background information and reduced clarity in rain images due to rain streaks and rain veiling effect,an image de-rain method based on deep frequency feature attention mechanism is proposed.Firstly,a parameter estimation network based on the frequency de-rain model is constructed,and a frequency feature decomposition module based on spatial scale transformation and octave convolution is designed to distinguish the frequency features of parameters,the spatial redundancy of low-frequency features is reduced,and the operation efficiency is improved.Secondly,in the rain streaks detection module,the multi-frequency channel attention mechanism is used to map the weight information of the rain streaks layers to enhance the weight feature diversity and improve the rain streaks detection performance.Finally,a de-rain repairing network based on focal frequency loss is designed to further repair the estimated image of the de-rain model,the image repair quality by narrowing the frequency gap is improved,and the anti-interference ability of the network is enhanced.The experimental results show that the proposed method can effectively remove the rain streaks in the real rain image scene,suppress the rain veiling effect,keep the background information of the image intact,and the image has high definition.

关 键 词：图像去雨 深度频率特征 多频通道注意力 倍频卷积 焦频损失 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391.4[电子电信—信息与通信工程]