WT-FFDNet：引入小波卷积的图像去噪网络
WT-FFDNet: Image Denoising Network with Wavelet Convolution

本文提出一种将小波变换引入FFDNet的图像去噪方法WT-FFDNet。WT-FFDNet通过结合深度可分离的小波变换卷积模块DepthwiseSeparableConvWithWTConv2d (DpSeWTConv)，利用多级小波分解提取图像的低频和高频特征，进行特征融合。网络进一步通过VGGBlock和结合通道注意力机制的ResNetBlock实现多尺度特征提取。在Berkeley Segmentation Dataset 500数据集上分别测试15、35、50、70噪声强度下的PSNR值，经过改进后的网络，其PSNR指标较原FFDNet网络有所提高，又在CBSD68、Urban100等数据集上测试不同噪声强度下的PSNR和SSIM指标，经过实验验证了该方法的有效性。
This paper proposes a wavelet-transform-integrated FFDNet method for image denoising, named WT-FFDNet. The WT-FFDNet incorporates a Depthwise Separable Convolution with Wavelet Transform module (DpSeWTConv) to achieve multi-level wavelet decomposition, extracting both low-frequency and high-frequency features of images for feature fusion. The network further implements multi-scale feature extraction through VGGBlocks and ResNetBlocks integrated with channel attention mechanisms. Experimental results on the Berkeley Segmentation Dataset 500 (BSD500) demonstrate improved PSNR metrics compared to the original FFDNet under noise levels of 15, 35, 50, and 70. Additional evaluations on datasets including CBSD68 and Urban100 across various noise intensities confirm enhancements in both PSNR and SSIM metrics, validating the effectiveness of the proposed method.