Four-channel convolutional Chinese handwriting recognition based on MobileNetV2

Abstract

Handwriting Chinese Character Recognition (HCCR) is the foundation of document digitization. It is a challenging subject in the field of image classification and recognition for a series of reasons such as the large number of Chinese characters, the diversification of writing style and numerous similar characters. To solve the above problems, this paper designs a four-channel convolution recognition model based on MobileNetV2. First, the input image is sent to four-channel convolution with different receptive fields, and feature maps of different scales are extracted respectively to improve the accuracy of the model. Then the feature maps are combined to enrich the diversity of features. Afterwards, the combined features are weighted by SE Block, and more useful feature maps are screened by this means to accelerate the model convergence. Finally, the lightweight network mobilenetv2 is used to classify the weighted features. The experimental results show that the recognition accuracy of the four-channel convolution recognition model based on mobilenetv2 on the offline handwritten Chinese character set CASIA-HWDB1.1 has reached 96.05%, and the convergence speed of the model is extremely fast. Also, the memory occupation and parameter quantity are far lower than other Chinese handwriting character recognition models.