基于改进MobileNetXt的轻量级手势识别方法
Lightweight Gesture Recognition Method Based on Improved MobileNetXt

轻量级卷积神经网络在复杂背景下的手势图像识别任务中存在网络结构单一、特征提取能力差、识别精度低等问题，本文针对上述问题提出一种基于改进MobileNetXt的轻量级手势图像识别方法。该方法在Sandglass沙漏块中引入CA注意力机制，捕捉跨通道的方向和位置感知信息；选用MetaAconC激活函数自适应地控制神经元的激活程度，增强模型的特征学习能力和梯度传播效率。该方法在公开的NUS-II手势数据集上识别准确率达99.12%，比原始的MobileNetXt网络提高了2.25%，与GhostNet和EfficientNet经典轻量级卷积神经网络相比，本模型在训练时间、模型参数量和识别准确率等网络指标上均有更优的性能表现。实验结果表明该方法有较强的信息表征和通道提取能力，在面临多种复杂背景因素干扰的手势识别挑战时仍具有较高的识别准确率。
Lightweight convolutional neural networks face problems such as a single network structure, poor feature extraction ability, and low recognition accuracy in gesture image recognition tasks in complex backgrounds. This paper proposes a lightweight gesture image recognition method based on improved MobileNetXt to address these issues. This method introduces a CA attention mechanism into the Sandglass hourglass block to capture cross channel direction and position perception information; The MetaAconC activation function is selected to adaptively control the activation level of neurons, enhancing the model’s feature learning ability and gradient propagation efficiency. This method has a recognition accuracy of 99.12% on the publicly available NUS-II gesture dataset, which is 2.25% higher than the original MobileNetXt network. Compared with GhostNet and EfficientNet classic lightweight convolutional neural networks, this model performs better in network metrics such as training time, model parameter count, and recognition accuracy. The experimental results show that this method has strong information representation and channel extraction capabilities, and still has high recognition accuracy in the face of gesture recognition challenges caused by various complex background factors.