Discriminative Deep Neural Network Model for In-air Handwritten Chinese Character Recognition

An end-to-end discriminative convolutional neural network (CNN) model is proposed to ex-tract discriminative features from the sequential coordinates of in-air handwritten Chinese characters, addressing the problem in existing CNN models that require data format conversion, thereby improving recognition speed. The Softmax cross-entropy loss function struggles to amplify inter-class differences. To address this, a discriminative loss function is introduced, which reduces the distance between a sample and the mean of its corresponding class while in-creasing the distance between the sample and the mean of its closest opposing class, thus enhancing the model's discriminative ability. Experimental results show that compared to the 1-dimensional CNN model, the proposed method achieves higher recognition accuracy and faster average recognition speed, with 96.90% accuracy and an average speed of 6 milliseconds.