Combination Of Convolutional Neural Network And Metric Learning For Offline Handwritten Chinese Character Recognition

Offline handwritten Chinese character recognition has always been the research focus and difficulty in the field of pattern recognition,because of the large amount of characters,multiple categories,complex structure of the character pattern,similar characters,distortion in the images,and so on.The traditional recognition framework “preprocessing+feature extraction+classifier” has not made great progress in recent years,and the deep learning method has injected new vitality into the field of compute vision.Therefore,more and more researchers try to apply deep learning methods to handwritten Chinese character recognition.At the moment,many commonly convolutional neural networks are widely used for handwritten Chinese character recognition,such as AlexNet,GoogLeNet,etc.,and have achieved better recognition results.However,the convolutional neural network only use cross-entropy as the loss function,and ignore the intra-class and inter-class distance of the input samples.Therefore,this paper proposes a new neural network combining a convolutional neural network with a metric learning module to increase the accuracy of classification task.The main contributions of this paper are as follows:Firstly,we propose to combine GoogLeNet with Triplet Loss and to combine ResNet with Triplet Loss to form two new network structures,and those two networks named GoogLeNet+Triplet Loss and ResNet+Triplet Loss.Those two networks are used for offline handwritten Chinese character recognition,and the recognition results of 90.07% and 97.07% are obtained.Triplet Loss module requires that the inputs must be triplet units,which two samples belongs to the same class and the other sample belongs to another class.The experimental results show that if the triplet units selection is unreasonable,it will seriously affect the recognition of the network.And only the reasonable Triplet Loss module can effectively reduce the intra-class distance of the input samples and expand the inter-class distance,and improve the final recognition results.Secondly,we propose to combine GoogLeNet with Center Loss and to combine ResNet with Center Loss to form two new network structures,and those two networks named GoogLeNet+Center Loss and ResNet+Center Loss.Those two networks are used for offline handwritten Chinese character recognition,and the recognition results of 96.18% and 97.03% are obtained.The results of experiments show that the Center Loss module can effectively improve the accuracy of offline handwritten Chinese character recognition.In addition,Center Loss has no the requirement of constructing triplet units,so it does not increase the computational complexity.Thirdly,we propose to combine ResNet with Center Loss using cosine distance for offline handwritten Chinese character recognition.This method effectively avoids the problem of losing the directional information of the input feature vectors when the traditional Center Loss is calculated using the Euclidean distance.We construct the new network consisting of ResNet and Modified Center Loss for offline handwritten Chinese character recognition,and obtain the accuracy of 97.24%.