MSBA-Net: Multiscale Behavior Analysis Network for Random Hand Gesture Authentication
Huilong Xie, Wenwei Song, Wenxiong Kang · IEEE Transactions on Instrumentation and Measurement · 2023
Random hand gesture authentication allows the probe hand gesture types to be inconsistent with the registered ones. While it is highly user-friendly, it poses a significant challenge that requires the authentication model to distill more abstract and complex identity features. Prior efforts on random hand gesture authentication mainly use convolution operations to obtain short-term behavioral information and cannot distill robust behavioral features well. In this paper, we propose a novel Multi-Scale Behavior Analysis Network (MSBA-Net), with a focus on capturing multi-scale behavioral features for random hand gesture authentication, which can simultaneously distill short-term behavioral information and model long-term behavioral relationships in addition to physiological features of hand gestures. In addition, as hand motion can result in inter-frame semantic misalignment, we propose an efficient semantic alignment strategy to mitigate this issue, which helps extract behavior features accurately and improves model performance. MSBA module is a plug-and-play module and could be integrated into existing 2D CNNs to yield a powerful video understanding model (MSBA-Net). Extensive experiments on the SCUT-DHGA dataset demonstrate that our MSBA-Net has compelling advantages over the other 20 state-of-the-art methods. The code is available at https://github.com/SCUT-BIP-Lab/MSBA-Net.