ID 2503: Spectrogram-Based Online Handwriting Recognition

Background

Current our project relies only on raw time-series IMU data for online handwriting recognition. However, handwriting motion contains patterns that may be more distinguishable in the frequency domain. By converting IMU signals into spectrograms, we capture both temporal and spectral features, enhancing pattern recognition. Spectrograms highlight periodic components, reduce noise, and reveal motion dynamics that raw signals may obscure. This approach has shown success in similar sequential data tasks, such as speech and bio signal analysis, suggesting its potential to improve handwriting recognition accuracy.

Tasks

• Modify current approach for spectrogram
  • Modify data processing
  • Implement both 1-D and 2-D backbones
  • Compare different implementations with similar numbers of parameters
• Support data collection

Requirements

• Proficiency in Python and PyTorch
• Knowledge in related field, e.g. time-series analysis, and speech recognition.
• Speak English and German to support data collection (optional)

Supervisors

Jindong Li

Researcher & PhD Candidate