FSMANet: Flash Shuffle Mix Attention Network for Human Sitting Posture Recognition
Tianxiang Zhao, Shoudong Shi, Kedi Qiu, Yunxin Ye, Ting Lan · 2024
Maintaining an incorrect sitting posture for a long time will cause health damage, so a sitting posture recognition system is needed to improve users’ sitting habits and prevent health problems. Recent findings have shown that lightweight convolutional neural networks exhibit excellent performance in a variety of computer vision tasks and significantly outperform other models in mobile deployments, making them valuable for practical applications. However, due to the locally-aware nature of the convolutional operations in convolutional neural networks, this leads to a limitation in their ability to capture long-range correlations, and thus a loss of sensitivity to critical information in the sitting discrimination task. To alleviate this dilemma, this paper proposes a new lightweight network architecture Flash Shuffle Mix Attention Network (FSMANet). First, we used our designed Flash Operation instead of regular convolution, which performs dynamic feature extraction on some of the channels of the input feature map according to specific rules, significantly reducing the redundancy in the feature map. Then, we introduce the Shuffle operation to rearrange the channels to enhance the information exchange between them. To overcome the problem of limited global information expressiveness, we propose a new 3D Mix Attention mechanism (MA), which enhances the informativeness and discriminability of feature descriptors by adaptively capturing long-range dependencies in all dimensions and extracting localized feature responses in different dimensions through three branches with almost no overhead. Experimental results show that FSMANet exhibits competitive performance in the sitting recognition task, achieving 96.86%(FSMANet) accuracy in our constructed high-quality dataset of 11 sitting categories, outperforming current state-of-the-art models.