Skew Correction of Handwritten Chinese Character Based on ResNet
Zetao Huang, Qian Zhang · 2019
Inclined characters can affect visual perception and optical character recognition (OCR). For the tilt problem of Chinese handwritten character, a 4-direction classification and a 181-angle classification model were proposed. These two models were built based on residual neural network (ResNet). After tuned and optimized, the models have implemented an end-to-end skew correction system. In the experiment of CASIA-HWDB1.1 data set, The Top-1 accuracy of the 4-classification model is 98.4% which is 3.1% higher than AlexNet, and the average loss of Top-1 score in the 181-classification model is 0.6° which is 1.7° shorter than AlexNet. Experimental results show that the model can effectively predict the tilt degree of Chinese handwriting character, also the corrected characters can obtain higher OCR recognition accuracy.