Handwritten New Tai Lue Character Recognition Using Convolutional Prior Features and Deep Variationally Sparse Gaussian Process Modeling

Abstract

New Tai Lue is widely used in Southwest China and Southeast Asia. Hence, it is important to study related handwritten character recognition. Considering the many similar characters in handwritten New Tai Lue, this paper proposes an offline handwritten New Tai Lue character recognition method based on convolutional prior features and deep variationally sparse Gaussian process (DVSGP) modeling. An offline handwritten database is constructed, a convolutional neural network is trained to extract the convolutional features of New Tai Lue character images as prior features, and a DVSGP model is built. The extracted features are input into the DVSGP model to construct a recognition model. The experimental results show that the accuracy of the model is 97.67% and that the precision, recall, and F1-score are 0.9769, 0.9767, and 0.9767, respectively, which are better than those of other methods. The proposed method also achieves high accuracy on the MNIST recognition task, verifying its universal applicability.