摘要
针对差分聚焦体(DFV)与变焦深度测量(DFF)联合应用的网络(简称为DFVDFF)精度较低的问题,首先,将特征提取模块的网络结构替换为UNet++,并在深度信息提取模块内增加注意力机制,通过关注重要的特征并融合深层特征和浅层特征提升网络预测的精度;然后,针对DFVDFF生成的深度图纹理边界模糊的问题,使用结构相似度和平滑平均绝对误差融合的损失函数替换原有的平滑平均绝对误差损失函数,通过提升损失函数对纹理边界的敏感程度引导网络生成更清晰的边界。实验结果表明,改进后的网络在有噪声的DDFF-12数据集上,相较于原始DFVDFF网络,均方误差下降了7.40%;在无噪声的FoD500数据集上,相较于原始DFVDFF网络,均方误差下降了19.07%。并且,改进后的网络在两个数据集上生成的深度图比DFVDFF网络生成的深度图具有更清晰的纹理边界。
In view of the low accuracy problem of the network jointly adopting DFV(Differential Focus Volume)and DFF(Depth From Focus)(referred to as DFVDFF),firstly,the network structure of the feature extraction module was replaced with UNet++,and an attention mechanism was added in the depth information extraction module to improve the accuracy of network prediction by focusing on important features and better fusing deep features and shallow features.Then,to address the problem of blurred texture boundaries in the depth map generated by DFVDFF,the loss function fused with structural similarity and smoothed mean absolute error was used to replace the original smoothed mean absolute error loss function to improve the sensitivity of the loss function to texture boundaries so that the network could be guided to generate clearer boundaries.Experimental results show that,compared to the original DFVDFF network,the mean square error of the improved network is reduced by 7.40%on noisy DDFF-12 dataset,and the mean square error is reduced by 19.07%on noise-free FoD500 dataset.Moreover,the depth maps generated by the improved network on both datasets have clearer texture boundaries than those generated by DFVDFF network.
作者
赵涂昊
夏小东
付茂栗
王觅
ZHAO Tuhao;XIA Xiaodong;FU Maoli;WANG Mi(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610041,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China;Shenzhen CBPM-KEXIN Banking Technology Company Limited,Shenzhen Guangdong 518206,China)
出处
《计算机应用》
CSCD
北大核心
2024年第S01期223-228,共6页
journal of Computer Applications
