Tracing Images with Neural Networks / Spring 2019 - Master Thesis
Many neural network models have been applied to generate images in a 2D/3D matrix format. The goal of this project is to build a model that is able to generate vector graphics. The envisioned approach is to train a model with reinforcement learning that can take bitmaps as input and output their vectorized representation. The vectorized representation is then turned into a bitmap and compared to the original to produce a quality score which is then used as the reinforcement signal.
The project would consist of designing a neural network model which can generate paths (polylines, curves, etc) as sequences of points. Additional steps include building an appropriate dataset with vector images of several orders of complexity (from simple shapes to more complex icons). Finally, the main step is building the neural network pipeline and training the model for the given task.
Supervised by Bojan Karlas
-  Xie, Ning, Hirotaka Hachiya, and Masashi Sugiyama. "Artist agent: A reinforcement learning approach to automatic stroke generation in oriental ink painting." IEICE TRANSACTIONS on Information and Systems 96, no. 5 (2013): 1134-1144. [arxiv]