ESRGAN-based visualization for large-scale volume data

Abstract: In the scientific visualization and data analysis workflow, data transmission bandwidth and memory resources have become the main bottlenecks when handling large-scale volume data. As a direct and effective scheme, data reduction is generally used to decrease data movement overhead and memory usage. However, it is still a challenge to obtain visualization results from reduced data without losing too many features. This paper proposes a visualization scheme for large-scale volume data based on enhanced super-resolution generative adversarial networks (ESRGAN) and designs a reduction-restoration workflow. Firstly, in order to reduce memory footprint, we propose an error-controlled data reduction method to delete data in 3D space, which is based on octree. Secondly, rendered images with loss of details are generated by performing volume rendering on reduced data. Lastly, to obtain feature-lossless visualization results, we apply ESRGAN to restore the details of rendered images. Based on the above scheme, the dual goals of data reduction and visual feature retention can be realized. Finally, the effectiveness of the proposed method is demonstrated by evaluating the performance of data reduction and visual restoration. Graphical abstract: [Figure not available: see fulltext.]