IRDC-Net: An Inception Network with a Residual Module and Dilated Convolution for Sign Language Recognition Based on...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Deaf and hearing-impaired individuals face communication challenges.
• Surface electromyography (sEMG) based sign language recognition (SLR) offers a promising solution for social integration.
• Traditional convolutional neural network (CNN) structures have limitations in capturing complex sEMG signal features.

Purpose of the Study:

• To propose a novel IRDC-net architecture for enhanced SLR.
• To improve the accuracy and efficiency of recognizing Chinese sign language signs.
• To validate the proposed method on a public dataset and compare it with existing CNN models.

Main Methods:

• Transformation of time-domain sEMG signals to the time-frequency domain using discrete Fourier transformation.
• Development and application of a novel Inception architecture with residual module and dilated convolution (IRDC-net) for SLR.
• Comparative analysis of IRDC-net against VGG-net and ResNet-18 using the Ninapro DB1 dataset.

Main Results:

• The IRDC-net achieved a classification accuracy of 91.70% on the custom Chinese sign language dataset after time-frequency transformation.
• On the public Ninapro DB1 dataset, the IRDC-net reached a classification accuracy of 89.82% for time-frequency data.
• The proposed IRDC-net outperformed VGG-net and ResNet-18 in SLR tasks.

Conclusions:

• The IRDC-net effectively captures intricate features from sEMG signals for improved SLR.
• Time-frequency domain transformation enhances the performance of sEMG-based SLR systems.
• This research contributes to advancing SLR technology and aiding deaf and hearing-impaired individuals.