| Abstract [eng] |
Creating cartoon characters is resource intensive and may usually draw out for months due to reasons such as: (1) prolonged generation of ideas, (2) stakeholders are not satisfied with the product or (3) legal procedures. This Master’s degree projects research phase sought to identify suitable quality metrics, evaluate existing solutions, and compared content generation technologies, but found a lack of solutions specific to the field of cartoon character generation. This project investigated the use of generative adversarial neural networks to automate the initial generation of the cartoon character, by converting sketches into raster images, as a potential solution to the resource intensive nature governed by the generation of ideas. The research hypotheses were that the quality of generated characters depends on (a) the quality of the input dataset and (b) by the size of the neural network. To test these hypotheses, eight different generative adversarial neural networks were trained on, respectively, a lower and a higher quality dataset. The experiments found a significant dependency on the quality of the dataset, but no distinguishable dependency on the size of the neural network. |
