In this thesis, we tackle the problem of automatic transfer function generation for volumetric data rendering. We study two methods using machine learning techniques. The first method involves gathering training data through suitable transfer function selection and classification by human users. We use these training transfer functions to optimize a generative neural network. With the second method we take an automated approach of generating volumetric data and labelling generated features, then training neural network by rendering volumes with generated transfer functions, and comparing the feature visibility on visualizations with expected render output. We compare both methods based on learning success and quality of generated transfer functions. The first method suffers from over-fitting due to small amount of training data, while with the second method we show that the training of the network cannot be performed using gradient descent method.
|
