分离门控与双分支条纹的自适应图像去雾算法  

Adaptive image dehazing algorithm with separable gating and dual-branch stripe

作　　者：吕伏 郭志昊 冯永安 LYU Fu;GUO Zhihao;FENG Yongan(School of Software,Liaoning University of Engineering Technology,Huludao 125105,China;Department of Basic Teaching,Liaoning University of Engineering Technology,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105 [2]辽宁工程技术大学基础教学部,辽宁葫芦岛125105

出　　处：《微电子学与计算机》2025年第4期58-69,共12页Microelectronics & Computer

基　　金：国家自然科学基金面上项目(51874166,52274206);国家自然基金青年基金(51904144)。

摘　　要：针对现有图像去雾算法存在的去雾不彻底、图像颜色失真和模型参数量大等问题,提出了一种分离门控与双分支条纹的自适应图像去雾算法。首先,在编码阶段重新设计分离卷积的门控线性残差块。通过优化分支结构,引入深度可分离卷积,有效地增强空间特征提取能力并减少了模型复杂度。其次,在特征转换阶段使用双分支条纹注意力模块。通过双分支整合水平和垂直局部条纹注意力,扩大感受野,从而保留重要的边缘特征。最后,在解码阶段构建自适应上下文感知融合模块。通过调整特征的空间权重,自适应地融合编码器部分低级特征与相应的高级特征,更好地恢复图像中的细节和纹理信息。实验结果表明,所提算法在SOTS室内数据集上仅以3.97 MB的模型参数量便使得峰值信噪比和结构相似性分别达到了36.77 dB和0.991。相较于传统去雾算法,所提算法细节保持完好,颜色自然,更符合人类的视觉感知。For the existing problems of incomplete defogging,color distortion,and large model parameter volume in existing image defogging algorithms,a novel adaptive image defogging algorithm with separated gating and dual-branch stripe is proposed.Firstly,in the encoding stage,the gated linear residual blocks with separated convolutions are redesigned,effectively enhancing spatial feature extraction capability and reducing model complexity by optimizing the branch structure and introducing depthwise separable convolutions.Secondly,in the feature transformation stage,a dual-branch stripe attention module is employed,integrating horizontal and vertical local stripe attention to expand the receptive field,thereby preserving important edge features.Finally,in the decoding stage,an adaptive context-aware fusion module is constructed to adaptively fuse low-level features from the encoder with corresponding high-level features by adjusting spatial weights of features,better restoring image details and texture information.Experimental results demonstrate that the proposed algorithm achieves PSNR and SSIM of 36.77 dB and 0.991 respectively on the SOTS indoor dataset with only 3.97 MB model parameters.Compared to traditional defogging algorithms,the proposed algorithm maintains intact details,natural colors,and is more consistent with human visual perception.

关 键 词：单幅图像去雾 门控线性单元 深度可分离卷积 注意力机制 特征融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]