结合迁移学习的真实图像去噪算法

Real-world Images Denoising Algorithm Combined with Transfer Learning

为了能有效地去除真实图像的复杂噪声,提出了一种结合迁移学习的真实图像去噪算法。该算法采用了双编码器结构,迁移学习编码单元利用预先训练好的权值有效提取鲁棒特征,残差编码单元对当前数据处理,进一步补充了信息。解码单元通过特征融合模块对丰富的信息进行融合,随后经过残差注意力模块加强对图像细节信息的关注,从而更好地恢复图像。实验结果表明,该算法在DND、SIDD和RNI15真实噪声数据集上有很好的泛化能力,能够在有效去除噪声的同时更好地保留图像纹理和边缘信息,恢复图像视觉效果更好。

To effectively remove the complex noise from real-world images,an image denoising algorithm combined with transfer learning was proposed.The algorithm was built with a dual encoder structure,in which the transfer learning coding unit effectively extracted robust features using pre-trained weights and the residual coding unit processed the current data.Combining the dual encoders further supplemented the information.The decoding unit fused the rich information by a feature fusion module and subsequently passed to a residual attention module to enhance the attention on image detail information so as to better restore images.Experimental results demonstrate that the algorithm achieves excellent generalization on DND,SIDD,and RNI15 real-world noise datasets.The algorithm can effectively remove noise while preserving image textures and edge information to restore images with better visual effects.