Offline Handwritten Chinese Character Recognition Based on Improved Googlenet

Abstract

Aiming at the problem of misrecognition in offline handwritten Chinese character recognition, this paper proposed an improved shallow GoogLeNet and an error elimination algorithm. Compared with the shallow GoogLeNet, the improved shallow GoogLeNet not only reduced the number of training parameters, but also maintained the depth of the Inception structure. According to the error elimination algorithm, the confidence of the samples in the test results was calculated and the erroneous samples in the dataset were removed. Then the dataset was divided into multiple similar character sets and one dissimilar character set. When the recognition result was in the dissimilar character set, it can be used as the final result. Otherwise, the final result could be obtained by the secondary recognition on the corresponding similar character set. The training and testing of the experiment were carried out on the CISIA-HWDB1.1 dataset. The accuracy of the method was 97.48%, which was 6.68% higher than that of the GoogLeNet network.