一种基于红外图像序列的深度学习三维重建仿真方法初探

Preliminary Study on a Deep Learning 3D Reconstruction Simulation Method Based on Infrared Image Sequence

为了提高红外图像的三维重建精度,提出一种基于红外图像序列、联合传统多视角恢复空间三维信息以及深度学习网络的三维重建仿真方法。该方法具体分为原始点云生成以及点云补全两部分。第一部分,通过输入预处理后的红外图像序列,基于传统被动式多视角三维重建方法生成不完整的原始点云;第二部分,将不完整点云通过一种联合式网络结构进行点云补全,进而实现三维重建。实验结果表明,本方法可以有效实现红外图像序列的点云生成,能够有效补全点云缺失部分,相比于现有点云补全方法,部分类别测试结果在搬土距离(earth mover’s distance,EMD)指标方面表现更好。此外,仿真模型具有良好的独立性以及拓展性,随着训练数据的增加以及训练方法的优化,具有高质量完成红外图像序列三维重建任务的潜力。

The paper proposes a 3 D reconstruction simulation method based on infrared image sequence.The method combined with traditional multi-view spatial 3 D information recovery and deep learning network, which is specifically divided into two parts: the raw point cloud generation, and point cloud completion. In the first part, the incomplete raw point cloud is generated based on the traditional passive multi-view 3 D reconstruction method by inputting the pre-processed infrared image sequence. In the second part, the incomplete point cloud is passed through a joint network structure, performing point cloud completion to a 3 D reconstruction. Experiments show that this method can effectively realize the point cloud generation of the infrared image sequence and can effectively restore the missing part of the point cloud. Compared with the existing point cloud completion methods, the test results of some categories performed better in EMD(Earth Mover’s Distance) index. With the increase of training data and optimization of training methods, it shows a potential to complete the 3 D reconstruction task of infrared image sequences with high quality.