Prediction and Generation of Multiple Complex Drawing Figures From Partial Drawing Sequences

Abstract

The goal of this study is to construct a model that predicts and generates the entire drawing sequence from a partial drawing sequence. In the proposed method, a recurrent neural network, namely Multiple Timescale Recurrent Neural Network (MTRNN), was used as the learning model. MTRNN has been modified to accommodate pen lifting. The experiment was performed using three functions of MTRNN (Learning, Recognition, Generation) and a drawing sequence consisting of the pen coordinates and the pen state. First, MTRNN training the drawing sequence and self-organizes the drawing dynamics. A partial drawing sequence is input to the trained MTRNN, and the recognition function calculates and predicts a vector that represents the entire drawing sequence. The entire drawing sequence is generated by inputting the calculated vector into the model. The results of the experiment were evaluated qualitatively, confirming the effectiveness of the proposed method.