Research On Painting-based Intelligent Transfer Function Design

DVR (Direct Volume Rendering) is one of the most popular techniques in volume data visualization. The internal essential information of volume data can be abstracted and shown interactively by DVR which provides the best way to observe internal structures of volume data. Volume data classification is one of the most important parts in DVR and is usually implemented by transfer function which assigns optical properties, such as color and opacity, to voxel and separates different kinds of voxels. However, it is very difficult to achieve satisfactory visualization results due to the large parameter space and unintuitive design process. Our research focus on intuitive, effective and intelligent transfer function design method design based on painting metaphor and machine learning. The details are as follows:(1) The Framework of Painting-based Interactive Transfer Function Design: In this framework, we employ painting information collected from painting-based user interface as training samples for ANN (Artificial Neutral Networks). Volume data are classified by the trained ANN and transfer function is generated automatically. Compared to traditional methods, painting-based user interface is more intuitive for transfer function design by allowing users to operate on images directly. After that, visualization could be achieved even if the intensity distribution of dataset is not available.(2) The Classification Strategy based on Statistical Properties: Local properties, such as intensity and gradient magnitude, are usually used as parameters for transfer function in traditional methods. In this scenario, the performance or classification are not as good as expected as local properties are sensitive to noise. Statistical properties of voxel are able to convey the distribution information of the same material of the voxel. Compared to local properties, statistical properties can be well exploited to depress noise and therefore, obtain better performance.(3) CUDA-Accelerated Interactive Transfer Function Design: As the training stage of ANN is compute-intensive and time consuming, ANN-based volume classification can not meet the interactive requirement of transfer function design. To achieve immediate feedback in painting process, ANN is accelerated by CUDA (Compute Unified Device Architecture) which is designed for applying GPU in high-performance computation.