Dense & Attention Convolutional Neural Networks for Toe Walking Recognition
Dense & Attention Convolutional Neural Networks for Toe Walking Recognition
复制标题
DOI：
10.1109/tnsre.2023.3272362
复制
发表时间：
2023-05
影响因子：
4.9
通讯作者：
Junde Chen;Rahul Soangra;M. Grant-Beuttler;Y. A. Nanehkaran;Yuxin Wen
中科院分区：
文献类型：
--
作者：
Junde Chen;Rahul Soangra;M. Grant-Beuttler;Y. A. Nanehkaran;Yuxin Wen
特发性脚趾行走（ITW）是一种步态障碍，儿童最初的接触显示有限或没有脚跟接触在步态周期。用脚趾走路会导致平衡能力差，增加跌倒或绊倒的风险，腿部疼痛，以及儿童发育迟缓。早期发现和识别有助于对诊断为ITW的儿童进行有针对性的干预。本研究提出了一种新的一维（1D）密集和注意力卷积网络架构，称为DANet，用于检测特发性脚趾行走。将密集块集成到网络中，最大限度地实现信息传递，避免遗漏特征。此外，注意力模块被整合到网络中，以突出有用的特征，同时抑制不需要的噪声。同时，增强了焦损功能，以减轻样品不平衡的问题。该方法优于其他方法，具有较好的性能。在从真实世界的实验场景中收集的本地数据集上识别特发性脚趾行走的测试召回率为88.91%。为了保证算法的可扩展性和泛化性，通过公开的数据集对算法进行进一步验证，所提方法的平均准确率、召回率和F1-Score分别达到89.34%、91.50%和92.04%。实验结果证明了该方法的有效性和可行性。
Idiopathic toe walking (ITW) is a gait disorder where children’s initial contacts show limited or no heel touch during the gait cycle. Toe walking can lead to poor balance, increased risk of falling or tripping, leg pain, and stunted growth in children. Early detection and identification can facilitate targeted interventions for children diagnosed with ITW. This study proposes a new one-dimensional (1D) Dense & Attention convolutional network architecture, which is termed as the DANet, to detect idiopathic toe walking. The dense block is integrated into the network to maximize information transfer and avoid missed features. Further, the attention modules are incorporated into the network to highlight useful features while suppressing unwanted noises. Also, the Focal Loss function is enhanced to alleviate the imbalance sample issue. The proposed approach outperforms other methods and obtains a superior performance. It achieves a test recall of 88.91% for recognizing idiopathic toe walking on the local dataset collected from real-world experimental scenarios. To ensure the scalability and generalizability of the proposed approach, the algorithm is further validated through the publicly available datasets, and the proposed approach achieves an average precision, recall, and F1-Score of 89.34%, 91.50%, and 92.04%, respectively. Experimental results present a competitive performance and demonstrate the validity and feasibility of the proposed approach.