摘要
针对真实世界图像去噪算法存在对上下文信息和全局信息利用不足导致的去噪效果不佳问题,提出一种U形金字塔注意力网络(UPCA)。U形结构由多尺度特征模块与长距离通道注意力模块融合形成的金字塔注意力模块组成,U形结构通过拼接操作可以将每一层的输出特征图融合,减少卷积过程以及下采样过程中图像细节特征的丢失。多尺度特征金字塔模块可以更好地利用上下文信息从而更好地恢复出干净的图像,而建立长距离依赖的通道注意力模块可以更好地利用全局信息,提高网络的去噪效果。同时在损失函数部分加入噪声项来加快训练时收敛的速度以及提高去噪效果。UPCA网络在数据集SIDD和DND进行对比实验,验证了UPCA网络的可行性和先进性,同时与同样使用通道注意力的RIDNet相比UPCA网络的PSNR/SSIM指标提升了0.81 dB/0.044,去噪后的效果图直观表现也更好,而且同等参数下训练所需的算力更小。
To address the issue of subpar denoising results in existing algorithms for real-world image denoising,we propose an innovative solution called the U-Shape Pyramid Channel Attention(UPCA).The U-shape structure comprises a fusion of multi-scale feature modules and long-range channel attention modules,forming a pyramid attention module.Through concatenation operations,the U-shape structure allows for the fusion of output feature maps from each layer,minimizing the loss of fine-grained image details during the convolution and downsampling processes.The multi-scale feature pyramid module effectively leverages contextual information to restore clean images,while the long-range channel attention module establishes dependencies on global information,thereby enhancing the denoising performance of the network.Additionally,we introduce a noise term in the loss function to expedite convergence during training and improve denoising efficiency.Experimental comparisons on the SIDD and DND datasets demonstrate the feasibility and superiority of the UPCA.Compared to RIDNet which also utilizes channel attention,UPCA achieves a remarkable improvement of 0.81 dB/0.044 in terms of PSNR/SSIM metrics.The visually enhanced denoised images produced by UPCA are superior,and it requires less computational power for training with the same set of parameters.
作者
王新武
陈春雨
WANG Xin-wu;CHEN Chun-yu(School of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)
出处
《计算机技术与发展》
2024年第4期48-54,共7页
Computer Technology and Development
基金
国家自然科学基金资助项目(61871142)
中央高校基本科研业务费项目(3072020CFT0803)。