Project Motivation and Description:
For my Compressed Sensing final project, my teammates and I examined the merits of modifying sparse recovery algorithms in order to improve classification of limb positions from electromyography (EMG) signals. In particular, we investigated performance in cases where the signal has degraded due to noise, causing classification accuracy to break down. Sparse Representation Classification (SRC) offered a promising approach for robust EMG classification with prior success in other noise-prone classification problems.
Our goal was to demonstrate how certain dictionary-manipulating methods may enhance already existing recovery techniques in both improving accuracy and reducing computation time.
My role:
We chose several methods to implement including dictionary augmentation, basis projection, dictionary compression, and sparse voting schemes. I focused ion dictionary augmentation with the training dataset collected from a single person. Mathematically speaking, dictionary augmentation expands the training dictionary dataset in attempt to help recover the uncorrupted EMG signal. Although dictionary augmentation slightly improved the classification of the limb position, it added extra computational cost from expanding the training dataset!
Most Challenging Part of the Project:
The most difficult part of this project was using limited amount of data to test these methods. While some conclusions were definitive–dictionary augmentation and basis projection neither improved the accuracy nor runtime, others were not. SRC voting and dictionary compression significantly cut down the runtime while maintaining the same level of accuracy as before. However, we would need more data to substantiate whether dictionary compression or SRC voting improves runtime and maintains the same level of accuracy for all subjects.
Further Work:
Our analysis contributed to Joseph Betthauser’s work and he used some of our code in his MATLAB package for EMG functions and classification methods for prosthesis control.
