Deep learning based feature extraction for handwriting styles clustering

Keywords: allograph clustering, autoencoder,clustering algorithms, deep learning,dynamic time warping, feature extraction, long-short term memory, machine learning, online handwriting, recurrent neural network.

A detailed analysis of existing methods for allograph clustering has been investigated and performed. We overviewed handcrafted (based on handcrafted feature extraction with further shallow clustering) and automated (NN-based) approaches. We proposed a new method for allograph clustering which combines learning efficient representations by RNN AE and DTW algorithm in a way that is not computationally expensive, but still provides a human-likeness of achieved results. We also evaluated and compared results based on internal metrics and pure visual analysis.

Korovai K. O. Deep learning based feature extraction for handwriting styles clustering : master’s thesis : 122 Computer science / Karyna O. Korovai. - Kyiv, 2021. - 64 p.