Handwritten Chinese text recognition method based on discriminative features

Abstract

The invention discloses a handwritten Chinese text recognition method based on discriminative features. The method comprises the following steps: firstly, preprocessing an input text image, including text correction and background cutting; then, a handwritten Chinese text recognition model based on CTC is constructed, the handwritten Chinese text recognition model comprises a feature extraction backbone network, auxiliary classifiers, a feature aggregation module and a main classifier, output probability prediction of all the auxiliary classifiers participates in loss calculation, the main classifier is combined with a center loss function to jointly optimize the model, and model optimization in a deep supervision form is achieved; and finally, reasoning prediction is carried out, and the handwritten Chinese text recognition model further improves the recognition precision by combining an N-gram language model and Beam Search decoding. According to the method, the performance of handwritten Chinese text recognition is remarkably improved by introducing a neural network intelligent algorithm including deep supervision, feature aggregation and center loss.

Images