基于结构引导Transformer的单视图三维重建去模糊方法

A Structure-Guided Transformer-Based Single-View 3D Reconstruction Deblurring Method

摘要: 随着XR与AR等交互式应用的迅速发展，利用图像进行三维重建在计算机视觉领域展现出重要价值。然而，实际拍摄图像过程中普遍存在的运动模糊会削弱纹理与结构信息，显著降低三维重建的几何一致性与细节完整度。为此，本文提出了一种面向单视图三维重建任务的结构引导Transformer去模糊网络。该方法引入了显式结构先验，通过结构引导前馈网络增强Transformer在模糊区域的边缘辨识能力；同时使用多头卷积自注意力模块降低传统自注意力的计算复杂度并加强局部空间建模能力。为了验证结构恢复对三维几何推断的有效性，本文将去模糊结果输入3D Gaussian Splatting的单视图重建框架中进行评估。实验结果显示，所提方法在多项指标上均取得更优表现。

Abstract: With the rapid development of interactive applications such as XR and AR, 3D reconstruction has demonstrated significant value in the field of computer vision. However, motion blur, which is prevalent in practice, weakens texture and structural information, significantly reducing the geometric consistency and detail integrity of 3D reconstruction. To address this, this paper proposes a structure-guided Transformer deblurring network for 3D reconstruction tasks. This method introduces an explicit structural prior and enhances the Transformer’s edge recognition ability in blurred regions through a structure-guided feedforward network; simultaneously, it uses a multi-head convolutional self-attention module to reduce the computational complexity of traditional self-attention and strengthen local spatial modeling capabilities. To verify the effectiveness of structural recovery for 3D geometric inference, the deblurring results are evaluated using a single-view reconstruction framework based on 3D Gaussian Splatting. Experimental results show that the proposed method achieves superior performance on multiple metrics.