融合精简双自适应注意力机制的图像复原算法  

Image restoration algorithm with integrated simplified dual adaptive attention mechanism

作　　者：王磊[1] 胡君红[1] 任洋 WANG Lei;HU Junhong;REN Yang(College of Physical Science and Technology,Central China Normal University,Wuhan 430079,China)

机构地区：[1]华中师范大学物理科学与技术学院,武汉430079

出　　处：《激光杂志》2025年第1期119-127,共9页Laser Journal

基　　金：国家自然科学基金(No.62101205);湖北省自然科学基金(No.2021CFB248)。

摘　　要：针对当前基于卷积神经网络的图像复原算法在处理物体运动场景时存在算法复杂度高、模型开销大、复原效果差等问题,提出了一种基于精简双自适应串行注意力机制的轻量化图像复原模型SCDNet。为降低模型复杂度,引入SimpleGate模块将特征图在通道维度上分成两部分并相乘以减少非线性激活函数带来的模型开销,采用精简双自适应串行注意力高效捕捉超像素级别的全局依赖关系,并自适应地传递到像素以提高算法对像素的表达能力,最后通过组合MS-SSIM和L1损失函数更好地保留图像的对比度、颜色和亮度等信息,提升了图像恢复质量。实验结果表明,SCDNet在GoPro数据集相对于Restormer算法PSNR提升0.30,SSIM提升0.12,而模型参数量仅为其22.4%。Addressing the issues of high algorithm complexity,large model overhead,and poor restoration performance in current convolutional neural network-based image restoration algorithms under object motion blur scenarios,we propose a lightweight image restoration model,SCDNet,based on a simplified dual self-adaptive serial attention mechanism.To reduce model complexity,we introduce the SimpleGate module,which splits feature maps into two parts along the channel dimension and multiplies them to reduce the model overhead caused by non-linear activation functions.We efficiently capture superpixel-level global dependencies using the simplified dual self-adaptive serial attention mechanism and adaptively transmit them to pixels to enhance the algorithm's pixel representation capability.Finally,by combining MS-SSIM and L1 loss functions,we better preserve image contrast,color,and brightness information,thereby improving image restoration quality.Experimental results show that,compared to the Restormer algorithm,SCDNet achieves a 0.30 increase in PSNR and a 0.12 increase in SSIM on the GoPro dataset,while its model parameters are only 22.4%of Restormer's.

关 键 词：双自适应 串行注意力 超像素 边缘细节 运动模糊 

分 类 号：TN249[电子电信—物理电子学]