基于注意力网络集成的联机空中手写识别研究

An Attention Network Ensemble-based Study of Online in Air-writing Recognition

针对联机空中手写识别的数据样本少、模型泛化能力不足、识别率低等问题,提出一种基于注意力网络集成的联机空中手写识别方法。该方法首先通过在形状特征中融入“联机”的时序特征,构建原始的多维数据;然后对多维融合数据降维投影到三个正交平面上,得到三组投影特征;其次,构建卷积神经网络用于提取视觉特征,同时引入字符嵌入作为图像的类标签,将类标签字符级语义特征通过注意力检测机制与三组视觉特征融合形成三组语义信息丰富的特征图,并基于特征图构建SoftMax分类器;最后,通过基于主学习器集成投票方法进行分类与识别。在两组空中手写数据集与哈工大(HIT-OR3C)联机数据上进行多组实验,在小样本的情况下,该方法识别率优于其他方法,分别达到95.68%,93.02%,94.96%。实验结果表明,该方法在小样本数据的情况下,充分发掘联机空中手写数据中有效特征,提高了空中手写识别效率。

Aiming at the problems of small data samples,insufficient model generalization ability and low recognition rate of online In Air-Writing recognition,an ensemble-based online In Air-Writing recognition method was proposed.Firstly,the original multi-dimensional data was constructed by incorporating“online”time series features into the shape features,and the multi-dimensional fusion data was projected to three orthogonal planes to obtain three sets of projection features.Secondly,a convolutional neural network was constructed to extract the visual features,next character embedding was introduced as class labels of the image,and the class-labelled character-level semantic features were fused with the three sets of visual features through the attention detection mechanism to form three sets of semantically informative feature maps,and a SoftMax classifier was constructed based on the feature maps.Finally,the classification and recognition was performed by the main learner-based integrated voting method.Multiple sets of experiments were carried out on two sets of In Air-Writing datasets and the HIT-OR3C online dataset,and in the case of small sample recognition,the recognition rates of the proposed method were better than that of other methods,which were 95.68%,93.02%and 94.96%respectively.The experimental result showed that the proposed method fully explored the effective features in the In Air-Writing data under the condition of small sample data,and improved the efficiency of In Air-Writing recognition.