Deep transfer learning-based adaptive gesture recognition of a soft e-skin patch with reduced training data and time

Deep learning-based classification algorithms are promising in gesture recognition with soft e-skin patches. However, the reported algorithms usually require large amount of training data, resulting in the time-consuming data collection process. In this paper, we present a deep transfer learning-based adaptive strategy for accurate gesture recognition of a soft e-skin patch with reduced training data and time. To this end, we first train a base neural network as the general feature extraction network. Next, we transfer the front layers of the pre-trained base network to target networks of new gesture recognition tasks. Further, we apply the fine-tune technique to refine the copied parameters. Finally, with our custom-built soft e-skin patch, we experimentally verify the developed strategy on two typical transfer cases, termed as the user transfer case (Case I) and the gesture transfer case (Case II). The experimental results show that, to ensure the stable accuracy of 95 %, the training data with and without the adaptive strategy are 1,312 vs 10,912 for Case I, and 8,192 vs 12,032 for Case II, respectively. In this sense, the training time of target networks can be reduced by 62.96 % for Case I and 34.20 % for Case II, respectively. This work shows the potential to promote the widespread application of e-skins in human computer interaction.