摘要： 地震数据的三维可视化能够直观反映地质的相关结构信息，为地质勘探等研究提供数据支持。针对传统体绘制算法在集中载入海量数据时存在显示延迟、画面跳跃、卡顿等问题，提出一种快速三维可视化算法。使用变分自编码器和深度聚类学习数据的空间特征表示，通过迭代优化目标函数提高聚类性能，以解决因空间数据分布不均造成的节点重叠问题。建立高效的索引结构，提高数据实时读取的效率，通过时序卷积网络预测下一个视点位置，提前将潜在数据载入内存，避免因海量数据集中加载导致画面卡顿、跳跃。使用基于双层视锥体的视点动态划分调度模型，剔除不必要的绘制节点及减轻系统负荷，从而提高数据渲染速度和流畅度。实验结果表明，该算法在索引结构上查询数据块的时间相比希尔伯特R树算法减少了64.14%~66.37%，预测视点的正确率相比拉格朗日插值算法提高了12.08%~22.70%，实时帧率在较大规模的子集上也能够相对稳定平滑，在保证图像质量的前提下整体系统的渲染性能达到预期效果。

关键词: 三维可视化, 深度学习, 深度聚类, 希尔伯特R树, 时序轨迹预测, 视锥体裁剪

Abstract: The 3D visualization technology of seismic data can intuitively analyze geological structure information, which can provide data support for geological exploration and other research.To solve the problems of display delay and image jump lag when loading large scale data in traditional volume-rendering algorithms, this study proposes a fast 3D visualization algorithm of massive seismic data based on deep clustering index.Firstly, the spatial feature representation of data is learned by variational autoencoding and deep clustering;then, the clustering performance is improved by iterative optimization of objective function.Thus, the problem of overlapping nodes caused by uneven spatial data distribution is solved, and an efficient index structure is established to improve the efficiency of real-time data reading.Additionally, a temporal convolutional network is used to predict the position of the next viewpoint, and the potential data is loaded into memory in advance to avoid the problem of picture-stuck jump caused by centralized loading when the data amount is too large.Finally, combined with the viewpoint dynamic-division scheduling model based on double-layer visual cone, redundant nodes are removed to reduce the system load, and improve the data-rendering speed and fluency.The experimental results show that, compared with the Hilbert R-Tree (HRT), the time of querying data sub-blocks on the index structure of algorithm in this paper is reduced by 64.14%~66.37%, and the accuracy of predicting viewpoints is improved by 12.08%~22.70%, compared with the Lagrange interpolation algorithm.The real-time frame rate can also be relatively stable and smooth on a large subset.The rendering performance of the whole system achieves the desired effect under the premise of ensuring the image quality.

Key words: 3D visualization, deep learning, deep clustering, Hilbert R-Tree(HRT), sequence trajectory prediction forecasting, view frustum culling

中图分类号: 

• TP391