HyFiNet: Hybrid feature attention network for hand gesture recognition

In this paper, we propose a portable CNN network: a hybrid feature attention network (HyFiNet) for precise hand gesture recognition. HyFiNet is designed by stacking four multi-scale refined edge extraction modules (REEMs). The REEM module is introduced to capture the refined edge information of hand gestures by incorporating hybrid feature attention (HyAttention) block. The HyAttention block is intended to focus on efficient salient features from multi-receptive fields and acquire knowledge of discriminable semantic structure for hand poses. As a resultant, multi-scale feature and hybrid feature attention mechanisms cohesively improve the performance of hand gesture recognition with a minimum computational cost. The efficiency of the proposed network is validated using six benchmark datasets: MUGD, Finger Spelling, NUS-I, NUS-II, HGR-I and Triesch, by adopting two validation schemes: person dependent and person independent. Furthermore, seven supplementary experiments are also performed for an ablation study to analyze the effectiveness of each module in the proposed network. The experimental results and visual representation indicate a substantial increase in accuracy compared to the existing state-of-the-art networks.