计算机科学
人工智能
质量评定
计算机视觉
特征(语言学)
可视化
传感器融合
模式识别(心理学)
特征提取
融合
质量(理念)
数据挖掘
数据可视化
图像融合
作者
Haoxue Gong,Haiyong Xu,Gangyi Jiang,Yeyao Chen,Ting Luo,Zhongjie Zhu
标识
DOI:10.1109/tvcg.2026.3738448
摘要
Novel view synthesis (NVS), including neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS), has demonstrated significant potential for 3D scene reconstruction and rendering, but issues of spatial fidelity and view consistency limit the quality of the generated images. Existing quality assessment methods predominantly rely on reference views, which limits their applicability when such views are sparse or unavailable. To address this issue, given that quality distortion in synthesized views arises from both spatial structural degradation in single images and geometric inconsistencies across views, we propose SVD-NQA, a no-reference quality assessment framework that jointly models spatial degradation and view-domain inconsistency in synthesized views. In the spatial domain, a multi-path perception network is constructed, leveraging three distinct models to capture local degradation, semantic context, and structural features, along with multi-scale feature fusion, cross-model feature interaction, and a high-frequency-guided sharpening (HFGS) module. In the view domain, we design inter-view difference maps via forward warping to capture drift artifacts, supplemented by higher-order singular value decomposition (HOSVD) for feature compression. Simultaneously, a bidirectional tensor unfolding slicing (BTUS) mechanism is introduced, locating distortion regions via difference maps and extracting feature slices that reflect the motion trajectories of geometric points. Finally, an autoencoder fuses the dual-domain features and regresses them to the subjective quality score. Experiments conducted on multiple NeRF and 3DGS datasets show that SVD-NQA significantly outperforms existing mainstream quality assessment methods in evaluation performance and exhibits reliable generalization capability.
科研通智能强力驱动
Strongly Powered by AbleSci AI