级联离散小波多频带分解注意力图像去噪方法  

Cascade discrete wavelet multi-band decompositionattention image denoising method

作　　者：王力 李小霞[1,2] 秦佳敏 朱贺[1,2] 周颖玥 Wang Li;Li Xiaoxia;Qin Jiamin;Zhu He;Zhou Yingyue(School of Information Engineering,Southwest University of Science&Technology,Mianyang Sichuan 621010,China;Sichuan Industrial Autonomous&Controllable Artificial Intelligence Engineering Technology Research Center,Mianyang Sichuan 621010,China;Sichuan Mianyang 404 Hospital,Mianyang Sichuan 621000,China)

机构地区：[1]西南科技大学信息工程学院,四川绵阳621010 [2]四川省工业自主可控人工智能工程技术研究中心,四川绵阳621010 [3]四川绵阳四0四医院,四川绵阳621000

出　　处：《计算机应用研究》2024年第1期288-295,共8页Application Research of Computers

基　　金：国家自然科学基金资助项目(62071399);四川省科技计划资助项目(2023YFG0262,2021YFG0383)。

摘　　要：针对图像去噪网络中下采样导致高频信息损失和细节保留能力差的问题,设计了一种级联离散小波多频带分解注意力图像去噪网络。其中多尺度级联离散小波变换结构将原始图像分解为多个尺度下的高低频子带来代替传统下采样,能减少高频信息损失。多频带特征增强模块使用不同尺度的卷积核并行处理高低频特征,在子网络每一级下重复使用两次,可增强全局和局部的关键特征信息。多频带分解注意力模块通过注意力评估纹理细节成分的重要性并加权不同频带的细节特征,有助于多频带特征增强模块更好地区分噪声和边缘细节。多频带选择特征融合模块融合多尺度多频带特征增强选择性特征,提高模型对于不同尺度噪声的去除能力。在SIDD和DND数据集上,所提方法的PSNR/SSIM指标分别达到了39.35 dB/0.918、39.72 dB/0.955。实验结果表明,该方法的性能优于主流去噪方法,同时具有更清晰的纹理细节和边缘等视觉效果。To address the issue of high-frequency information loss and poor detail preservation ability in image denoising networks caused by downsampling,this paper proposed a cascade discrete wavelet multi-band decomposition attention image denoising network.The multi-scale cascade discrete wavelet transform structure decomposed the original image into high and low-frequency sub-bands at multiple scales,replacing traditional downsampling and reducing high-frequency information loss.The multi-band feature enhancement module employed convolutional kernels of different scales to process high and low-frequency features in parallel.By repeating this process twice at each level of the subnetwork,it effectively enhanced both global and local key feature information.The multi-band decomposition attention module evaluated the importance of texture detail components through attention and weighted the detail features of different bands,which helped the multi-band feature enhancement module better distinguish between noise and edge details.The multi-band selective feature fusion module fused multi-scale multi-band features to enhance selective features,improving the model’s ability to remove noise at different scales.The proposed method achieves PSNR/SSIM values of 39.35 dB/0.918 and 39.72 dB/0.955 on the SIDD and DND datasets,respectively.The experimental results demonstrate that the proposed method outperforms mainstream denoising methods and produces clearer visual effects,such as texture details and edges.

关 键 词：图像去噪 高频信息 级联离散小波变换 多频带特征增强 多频带分解注意力 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]