Related Experiment Videos
A model-based sequence similarity with application to handwritten word spotting
José A Rodríguez-Serrano1, Florent Perronnin
1Textual and Visual Pattern Analysis Group, Xerox Research Centre Europe, Meylan, France. Jose-Antonio.Rodriguez@xrce.xerox.com
IEEE Transactions on Pattern Analysis and Machine Intelligence
|January 18, 2012
Summary
This study introduces a new method for comparing vector sequences using semicontinuous Hidden Markov Models (SC-HMMs). This approach improves accuracy and significantly reduces computational costs for tasks like handwritten word retrieval.
Area of Science:
- Pattern Recognition
- Machine Learning
- Sequence Analysis
Background:
- Model-based approaches are common for sequence similarity.
- Hidden Markov Models (HMMs) are widely used for sequence modeling.
- Traditional methods like Dynamic Time Warping (DTW) can be computationally expensive.
Purpose of the Study:
- To propose a novel similarity measure for vector sequences.
- To introduce semicontinuous Hidden Markov Models (SC-HMMs) for sequence modeling.
- To reduce the computational cost of sequence similarity computation.
Main Methods:
- Sequences are mapped to semicontinuous Hidden Markov Models (SC-HMMs).
- SC-HMMs utilize mixtures of shared Gaussians for emission probabilities.
- Similarity is computed using Dynamic Time Warping (DTW) on mixture weight vectors.
Main Results:
- The proposed SC-HMM similarity measure outperforms traditional DTW and HMM-based methods.
- Experiments on handwritten word retrieval datasets demonstrate improved accuracy.
- A significant reduction in computational cost is achieved.
Conclusions:
- SC-HMMs offer a more accurate and computationally efficient approach to sequence similarity.
- The method is effective for handwritten word retrieval tasks.
- The trade-off between accuracy and computational cost can be managed.