高频信息物体多层多元特征权重自适应融合三维重建网络

打开文本图片集

关键词：深度学习；光度立体视觉；多元卷积；特征融合；自适应权重中图分类号：TP394.1;TP183 文献标识码：Adoi：10.37188/OPE.20253315.2424 CSTR：32169.14.OPE.20253315.2424

Multi-layer multi-feature adaptive weight fusion network for 3D reconstruction of objects with high-frequency information

WANG Biao，LI Ying，RONG Baichuan，LIU Jing，ZHANG Jin，WANG Yonghong （School of Instrument Science and Opto-electronics Engineering，Hefei University of Technology，Hefei230000，China） * Corresponding author，E-mail： yhwang@hfut. edu. cn

Abstract：To mitigatethe loss of high-frequency surface texture information and the resulting reduction in reconstruction accuracy in deep learning-based photometric stereo，a multi-layer multi-element feature weight adaptive fusion 3D reconstruction network （MMF-Net） is proposed.The network architecture builds on PS-FCN and incorporates a symmetric encoder-decoder to enhance feature learning，representationcapacity，and multi-level feature integration.A novel multi-element convolution layer with independent inter-layer adaptive weights is introduced；by incorporating additional trainable weights，both shape and texture cues are jointly leveraged to better capture fine surface texture variations，thereby improving stabilityand accuracy in scenes containing dense high-frequency information.An auxiliary skip-connection mechanism is also employed to propagate intermediate-layer features to later stages，preserving high-frequency details while reinforcing low-frequency structure，and enabling efective fusion of multi-band （highand low-frequency） surface information. The method was evaluated on the DiLiGenT benchmark. MMFNet attains an average mean angular error（MAE）of 6.94∘ ，representing a 6% improvement over PSFCN（Norm）at 7.39∘ . For objects exhibiting pronounced high-frequency surface detail， the average reconstruction error is 11.03∘ ，a 12% improvement relative to FUPS-Net at 12.52∘ . The results demonstrate that MMF-Net efectively captures both low- and high-frequency surface information in photometric stereo，offering a viable approach for high-precision 3D reconstruction from surface normals.

Key words： deep learning；photometric stereo vision；multivariate convolution； feature fusion；adaptive weighting

1引言

物体表面法向量的准确获取在计算机视觉、三维重建、机器人视觉以及虚拟现实等领域具有极其重要的应用价值。（剩余25338字）