基于编解码多尺度特征优化的图像去雾算法  

Image dehazing algorithm based on multi-scale feature optimization in encoding and decoding

作　　者：邵小桃[1] 郭燕 申艳[1] 钱满义[1] SHAO Xiaotao;GUO Yan;SHEN Yan;QIAN Manyi(School of Electronic and Information Engineering,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学电子信息工程学院,北京100044

出　　处：《北京交通大学学报》2024年第2期37-46,56,共11页JOURNAL OF BEIJING JIAOTONG UNIVERSITY

基　　金：装备预研教育部联合基金(8091B0203)。

摘　　要：真实雾气不均匀分布的特点会导致基于合成数据集训练的网络对真实雾气下拍摄的图像的复原质量不佳.此外,现有去雾模型较大的网络参数量会影响去雾的实时性.针对这两个问题,提出一种参数量较低的基于编解码多尺度特征优化的图像去雾算法以去除真实场景下图像的雾气.首先,在编码部分利用跨通道上下文注意力隐式地建模像素间的关系,以恢复去雾后图像中物体的结构.然后,设计信息调节子网弥补编码器遗漏的浅层信息,解决细节恢复粗糙的问题.最后,在解码部分设计特征矫正子网,采用相减式残差结构减少噪声,保证输出结果的正确性.在多种真实雾数据集上,对所提方法的普适性进行实验.实验结果表明:在REVIDE真实雾数据集中,与MSBDN方法相比,所提方法在参数量降低46%的基础上获得了PSNR 1.25dB的提升;在OHaze、I-Haze以及RTTS多种室内外真实雾测试集中,与其他去雾方法相比,所提方法都取得了更好的PSNR结果和视觉效果.The uneven distribution of real-world haze often leads to poor recovery quality of images captured under actual haze conditions by networks trained on synthetic datasets.Additionally,the large number of parameters in existing dehazing models affects the real-time performance of haze removal.To address these two issues,this paper proposes an image dehazing algorithm based on multi-scale feature optimization in encoding and decoding with a reduced number of parameters to ef-fectively remove haze from real-world images.Firstly,the encoding part employs cross-channel con-textual attention to implicitly model the relationships between pixels,thereby preserving the structure of objects in the dehazed images.Then,an information tuning subnet is designed to compensate for the shallow information missed by the encoder,addressing the problem of coarse detail recovery.Finally,the decoding part incorporates a feature correction subnet that utilizes a subtractive residual structure to reduce noise and ensure the accuracy of the output.Experiments on various real-world haze datasets demonstrate the generalizability of the proposed method.The results show that on the REVIDE real-world haze dataset,the proposed method achieves a 1.25 dB improvement in PSNR compared to the MSBDN method,with a 46%reduction in the number of parameters.Furthermore,on the O-Haze,I-Haze,and RTTS real-world indoor and outdoor haze test sets,the proposed method outperforms other dehazing methods in terms of PSNR and visual quality.

关 键 词：信号与信息处理 图像去雾 深度学习 真实雾 编解码 

分 类 号：TN911.73[电子电信—通信与信息系统]