计算机工程

• •    

基于2D2DPCA-CNN的美式手语识别研究

  

  • 发布日期:2020-12-30

Research on American Sign Language Recognition Based on 2D2DPCA-CNN

  • Published:2020-12-30

摘要: 针对美式手语(American Sign Language,ASL)识别时存在识别准确率较低,模型训练时间过长等问题,提出了一种 双向二维主成分分析(2D2DPCA)与卷积神经网络(CNN)相结合,使用贝叶斯算法(Bayesian Optimization,BO)优化模型参数 的算法。该算法首先使用 2D2DPCA 算法对原始图片数据降维,提取行、列方向的特征图;然后使用卷积神经网络对特征图 进行训练分类;最后使用贝叶斯优化对模型超参数进行自动调参。在 24 分类 ASL 数据集上,该算法达到 99.15%的识别准确 率,运行时间相比未经过预处理的算法提升 10.3 倍。结果表明该算法在 ASL 识别中,准确率更高,所耗时间更短。

Abstract: In order to solve the problems of low recognition accuracy and long model training time in American Sign Language (ASL) recognition, this paper proposed an algorithm that combines bidirectional two-dimensional principal component analysis (2D2DPCA) and convolutional neural network (CNN) and uses Bayesian Optimization (BO) to optimize model parameters. Firstly, the algorithm used the 2D2DPCA algorithm to reduce the dimensionality of the original image data, and extracted the feature maps in the row and column directions; then used the convolutional neural network to train and classify the feature maps; and finally used Bayesian optimization to adjust the model hyperparameters automatically. On 24 classified ASL data sets, the algorithm achieved a recognition accuracy of 99.15%, and the running time was 10.3 times faster than the algorithm without preprocessing. The results show that the algorithm has higher accuracy and shorter time in ASL recognition.