Table 1 Hyper-parameter tuning on DB-a.

Baseline

Without any preprocessing or data augmentation. Create a RGB image by replicating grayscale sEMG image three times. Use the same ConvNet architecture as that in experiment 1, with the exception that the last two convolutional layers are removed. The recognition accuracy reached 92.5% using simple majority voting over 6 frames.

Malfunctioning channels corrected

Detect and correct malfunctioning channels as described in a previous study⁴⁵.

Power-line interference removed

Band-stop filtered between 45 and 55 Hz using a second-order Butterworth filter to remove power-line interference.

Full-wave rectified

Full-wave rectification followed by low-pass filtering is a standard amplitude estimation technique⁴⁶. Full-wave rectification takes the absolute value of the sEMG signals.

Low-pass filtered

Full-wave rectified and low-pass filtered using a second-order Butterworth filter with a cut-off frequency of 75 Hz.

Amplitude normalized

sEMG signals are normalized by maximum voluntary contraction (MVC), which is a technique used to reduce the unwanted variability by dividing the sEMG signals by a reference value¹⁰. For each subject, the reference value is computed as the maximum of the full-wave rectified and low-pass filtered (3 Hz, second-order Butterworth) max-force data which is collected during the same acquisition session with eight gestures.

Cross-talk removed

sEMG signals are preprocessed by independent component analysis (ICA), which has been found to be successful for cross-talk removal⁴⁷.

Jet colour scheme

Create a RGB image from grayscale sEMG image by Jet colour scheme⁴⁸.

Training data augmented

Augment training data by circularly translate ±1 pixels in row direction to simulate electrode shift⁴⁹.