基于注意力机制的多级小波CNN遥感图像去噪算法  

Remote Sensing Image Denoising Based On Multi-Level Wavelet CNN with Attention Mechanism

作　　者：成丽波[1] 苑浩然 李喆 贾小宁 

机构地区：[1]长春理工大学数学与统计学院,吉林 长春

出　　处：《计算机科学与应用》2024年第4期73-82,共10页Computer Science and Application

摘　　要：高质量的卫星遥感图像对计算机视觉任务来说尤为重要,在去噪模型中,感受野大小和效率之间的权衡是去噪任务的一个关键问题。普通卷积网络(CNN)通常以牺牲计算成本为代价来扩大感受野。通道注意力机制可以保证去噪性能的同时尽量减少计算成本,本文中提出了一种新的基于注意力机制的多级小波CNN模型,以更好地平衡感受野大小和计算效率。在改进U-Net结构的基础上,引入小波变换来减小收缩子网络中特征图的大小。此外,通过通道注意力机制进一步优化模型,使对噪声成分的提取更加有针对性。实验采用峰值信噪比(PSNR)和结构相似性(SSIM)两项评价指标对实验结果进行量化评判,在高斯噪声标准差为15,25,30时,较DNCNN,FFDNET等方法在PSNR值上平均提高10%左右,图像细节清晰,能有效地保护遥感图像边缘特征。High quality satellite remote sensing images are particularly important for computer vision tasks, and the trade-off between receptive field size and efficiency is a key issue in denoising models. Ordinary convolutional network (CNN) typically sacrifice computational costs to expand the receptive field. The channel attention mechanism can ensure denoising performance while minimizing computational costs. In this paper, a new multi-level wavelet CNN model based on attention mechanism is proposed to better balance receptive field size and computational efficiency. On the basis of improving the U-Net structure, wavelet transform is introduced to reduce the size of feature maps in the shrinking sub network. In addition, the model is further optimized through channel attention mechanism to make the extraction of noise components more targeted. The experiment uses two evaluation indicators, peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), to quantitatively evaluate the experimental results, when the standard deviation of Gaussian noise is 15, 25, and 30, compared to methods such as DNCNN and FFDNET, the average PSNR value is improved by about 10%, and the image details are clear, which can effectively protect the edge features of remote sensing images.

关 键 词：深度学习 图像去噪 卫星遥感图像 小波变换 注意力机制 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]