一种基于两步法的时空域图像增强算法及其在三维重建中的应用

A Two-step Spatiotemporal Based Image Enhancement Algorithm and Its Application in 3D Reconstruction

为了提高基于多视图立体匹配网络(MVSNet)的深度学习算法在复杂场景下的重建完整性,提出了一种基于两步法的时空域图像增强算法,可用于提高视频图像的质量,并最终提高三维模型的重建完整性。首先利用空间域去噪网络减少视频帧内噪声,在这一步有效地降低噪声水平,使得下一步能够更好地处理帧间噪声和建立网络模型。在第二步时间域视频增强网络利用相邻帧间存在的时间信息,能够有效地减少帧间闪烁。使用公开数据集Tanks and Temples中的Truck视频图像进行验证,实验结果证明,在使用同一个多视图立体匹配网络R-MVSNet下生成三维点云,相较目前先进方法FastDVDnet,该方法误匹配点数更少,点云更完整、更清晰。

This paper proposes a two-step spatiotemporal based image enhancement algorithm that can be used to improve the quality of images and ultimately the reconstruction completeness of 3D models.Firstly,it uses spatial domain image denoising network to reduce the intra-frame noise,this stage can effectively reduces the noise level and enables the next step to better handle inter-frame noise and build the network model.In the second step,this paper uses the time-domain image enhancement network to effectively reduce flickering by the temporal information that exists between adjacent frames,and uses Truck video images from the publicly available dataset Tanks and Temples to validate this two-step spatiotemporal network.The experimental results demonstrate that when using the same multi-view stereo matching network R-MVSNet to generate 3D point clouds,this method has fewer false matches and more complete and clear point clouds than the current state-of-the-art method FastDVDnet.