结合高低频分解和多尺度两级融合策略的单一深度图去噪方法

Single depth map denoising method by combining high-and low-frequency decomposition and multi-scale two-level fusion strategy

现有的很多基于神经网络的深度图去噪方法忽略了深度图结构信息和细节信息之间的差异性,导致这些方法往往无法准确地恢复深度图的几何结构信息.为此,提出一种结合图像高低频分解和多尺度两级融合策略的单一深度图去噪方法.首先,考虑到不同噪声的差异性,引入多尺度高斯滤波器将含噪声的深度图分解为一组低频结构分量和一组高频细节分量.其次,考虑到低频结构信息和高频细节信息的互补特性,将这两组分量分别输入到基于多尺度两级融合的低频特征提取网络和高频特征提取网络,并且提出在这两个网络之间使用互补的特征加权融合机制进行多级特征融合和反馈.最后,对高低频特征提取网络输出的高低频增强特征,使用高低频合并重建模块进行残差预测,再将其与输入图相融合得到高质量的深度图.实验结果表明,在峰值信噪比、均方根误差、结构相似性和综合性能对比方面,本文方法比多个主流的深度图去噪方法如ARCNN、Fast ARCNN、DnCNN、ADNet和FFDNet拥有更好的性能.

Many existing neural network-based depth map denoising methods ignore the difference be‐:tween depth map structure information and detail information,leading to the fact that these methods often fail to accurately recover the geometric structure information of the depth map.To this end,this paper proposes a single depth map denoising method that combines image high-frequency(HF)and low-frequency(LF)decomposition and multi-scale two-level fusion(MTF)strategies.First,consider‐ing the difference between different noises,a multi-scale Gaussian filter is introduced to decompose the depth map containing noise into a group of LF structural components and a group of HF detail com‐ponents.Secondly,considering the complementary characteristics of LF structure information and HF detail information,these two groups of components are input to the LF feature extraction network(LFN)and HF feature extraction network(HFN)based on MTF,and a multi-stage feature fusion and feedback is proposed using a complementary feature weighting fusion mechanism between these two networks.Finally,the HF and LF enhanced features output from the HFN and LFN are predicted us‐ing the high-and low-frequency merged reconstruction module for residuals,and then fused with the input map to obtain a high-quality depth map.The experimental results show that the proposed method has better performance than several mainstream deep graph denoising methods such as ARCNN,Fast ARCNN,DnCNN,ADNet,and FFDNet in terms of peak signal-to-noise ratio,root mean square error,structural similarity,and comprehensive performance comparison.