水下折射序列图像非单视点相机模型的深海场景自动重建研究

3D reconstruction for underwater scene applied in ocean exploration draws researchers’ attention in recent years. The water absorbs, scatters and refracts light, which makes the color images and video sequences collected underwater blur and noisy. Therefore, the 3D Reconstruction methods on land cannot be directly transplanted into underwater scene. This project designs a framework of 3D reconstruction for underwater deep-sea scene. The main contributions of this project will be as follows: .We will present a new quaternion vector filter for removal of random impulse noise in color images or video sequences from underwater scene to achieve noticeable performance improvements. .The virtual camera error function will be introduced. It builds upon the idea where a virtual camera is defined for each 2D point into which the corresponding 3D point can be projected. We also propose a new underwater refractive bundle adjustment, for which the virtual camera error function is an implicit constraint..Due to the refractive camera's characteristics it is not possible to directly apply epipolar geometry or rectification to mages because the single-view-point model and, consequently, homographies are invalid. Additionally, the projection of 3D points into the camera cannot be computed efficiently. Therefore, the method proposed is an adapted plane sweep algorithm that is based on the idea of back-projecting rays for each pixel and view onto the 3D-hypothesis planes..We will set up an actual automated 3D reconstruction approach, which follows a common sequential SFM and for each key frame compute dense distance information per pixel (a depth map) and fuse the depth maps to obtain the final 3D model..These methods and algorithms will be proved and evaluated in a set of experiments, and the underwater scene 3D reconstruction system will be applied into other projects in practice.

本项目针对海洋开发行业急需的深海水下目标及环境的三维重建关键技术进行研究，旨在建立快速、准确、自动的水下三维重建方法。基于水下光学图像的三维重建是当前水下视觉领域的研究热点问题之一，由于海洋开发环境中的各种矿砂、杂质、气泡、浮游生物的存在，水下光照不足，以及光线的散射和折射作用，使水下三维重建面临困难重重。本项目重点解决水下彩色图像去噪、水下折射引发系统模型误差导致的三维重建失败的问题，并且有如下创新：提出了一种基于四元数表达的水下彩色图像脉冲去噪方法；引入了水下虚拟相机误差函数，并提出明确考虑折射的水下重建优化策略以及基于虚拟相机反向投影的稠密深度图快速计算的方法。本课题将建立一种明确考虑折射引入的系统误差的通用自动水下三维重建方法，还将结合本单位的实际工程项目来验证课题中所设计的原型系统和各项关键算法在解决水下三维重建面临的问题的正确性和有效性。

在海洋开发水下作业中，基于图像的水下三维重建技术是热点研究问题之一。然而，由于海洋开发环境中的各种矿砂、杂质、气泡、浮游生物的存在，同时还存在光线在水中的衰减和散射，采集到的原始水下图像往往存在色差、噪声和分辨率低等问题，并且，光线的折射会给三维重建带来较大误差。本项目针对以上问题，开展了水下图像增强算法以及考虑折射的水下三维稠密重建算法研究。研究了基于神经网络的水下图像增强算法以及超分辨率算法，提出一种能同时实现水下图像增强和超分辨率的方法，使用SFM方法（图像处理方式）随机预先处理水下图像，将生成对抗网络融合水下白平衡图像特征，生成器采用了GhostNet网络结构，该网络结构能够大量减少计算资源，提升计算速度，提高图像超分辨率质量，实验证明该方法能够同时实现水下图像的增强以及超分辨率并且计算量更小，能够同时完成对水下图像进行颜色校正、去噪、细节恢复。本项目同时研究了考虑折射的三维稠密重建方法，提出了计算折射的虚拟相机成像模型，提高了位姿计算的精确度，结合计算的深度图，进行了稠密重建，实验证明，该方法能够提供准确的水下稠密重建，并且能够给出绝对度量值。本项目的研究成果能够将水下目标进行自动、快速、准确的三维重建，能够为已经被广泛应用于海洋生物检测与水下设施检查的水下机器人提供准确的水下目标形态数据，指导水下作业以及做出分析与决策。