Abstract:Neural implicit surface reconstruction has achieved significant progress in recovering detailed 3D shapes from images. However, existing methods predominantly rely on multi-layer perceptrons (MLP) for scene encoding, which lack explicit 3D structural information, thus limiting the reconstruction quality in terms of geometric detail and global consistency. To address this limitation, this study proposes a hierarchical volume-based encoding framework that explicitly embeds spatial information into neural implicit representations. Specifically, a multi-scale volume encoding strategy is introduced, in which high-resolution voxels capture spatially-varying features to preserve high-frequency geometric details, while low-resolution voxels maintain spatial consistency and shape smoothness by sharing features across neighboring positions. To improve memory efficiency, sparse structures are adopted to reduce the overhead introduced by high-resolution voxels, and two regularization terms are incorporated to further enhance the smoothness of the reconstructed surfaces. The proposed voxel encoding module is designed as a plug-and-play component, allowing seamless integration into various neural implicit surface reconstruction frameworks. Experimental results on DTU, EPFL, and BlendedMVS benchmark datasets demonstrate that this method achieves substantial improvements across multiple evaluation metrics, producing high-quality reconstructions that combine smooth surfaces with rich geometric details.