Abstract by Taylor Archibald
Scan, Attend, Recover: A Deep Encoder-Decoder Network to Recover Handwritten Strokes from Images
Stroke order and velocity are helpful features in the fields of signature verification, handwriting recognition, and handwriting synthesis. We use a deep neural network (DNN) encoder-decoder architecture with attention to infer temporal stroke information from static images of handwritten text. We believe this is the first DNN approach to stroke recovery that accommodates variable width inputs. We further propose several loss functions that are novel for this task. Finally, we demonstrate that our recovered temporal stroke information can improve both handwriting recognition and handwriting synthesis.