Research On End-to-end Offline Handwriting Recognition Technology Based On Deep Learning

Off-line handwriting recognition is a process of converting images containing handwritten characters into text information that can be edited by the computer.It is widely used in processing historical documents,e-mail information,administrative documents and other fields.Off-line handwriting recognition equipped with the "fast train" of deep learning is an important branch of research in the field of pattern recognition and artificial intelligence.At present,the end-to-end model based on deep learning,which solves the problems of cumbersome recognition process of traditional methods and difficult to design models with good generalization performance,has gradually became a research hotspot in the field of off-line handwriting recognition.The end-to-end off-line handwriting recognition can accomplish character detection and recognition at the same time,which greatly improvs the real-time character recognition.Due to the freedom of the writing,different instances of the same character,different people will inevitably appear a series of quite different glyphs.In the same sense,the same character written by the same people,due to the different writing context environment,will be great differences in writing.At present,the research on off-line handwriting recognition is mainly aimed at databases with small amount of data.It is very expensive to quickly collect huge training set data with annotation.The diversity of writing style and the scarcity of labeled data make off-line handwritten recognition still a challenging subject.The traditional off-line handwriting recognition process includes image preprocessing,character segmentation,feature extraction and recognition.At present,the more popular method end-to-end handwriting recognition based on deep learning,which converts handwritten text into a sequence of recognition problem,and realizes the mapping of image sequence to character sequence through sequence-to-sequence model.This paper studies the problem of off-line handwriting recognition,and based on the end-to-end model structure of by using deep learning,the following research work is carried out on off-line handwriting recognition:End-to-end offline handwriting recognition based on deep learning currently mainly has two models based on Connection Timing Classification and based on Attention mechanism,which have good recognition accuracy.However,the CTC-based model assumes that there is conditional independence between tags.Each time the output is an independent single character probability,ignoring the overall information;The attention-based model does not introduce any constraint conditions for guiding alignment,and it is prone to dislocation for handwriting text recognition with strict monotonic input and output alignment.In view of the above problems,this paper proposes an end-to-end off-line handwriting recognition model based on CTC-Attention mechanism.Based on the Encoder-Decoder framework,it uses multi-task learning and introduces a method of dynamically adjusting parameters to balance the CTC model and the Attention mechanism model.The effectiveness of the model is verified on the open source offline handwritten English word IAM dataset and offline handwritten Chinese text line CASIA-HWDB2 dataset.Experiments show that the proposed model has a higher recognition accuracy.Although the hybrid model proposed in(1)can effectively realize end-to-end offline handwriting recognition,such a hybrid model has a complex structure and takes a long time to train.To train a model with better performance,a large amount of labeled data is required.At present,the public dataset size of off-line handwriting recognition cannot reach the data scale comparable to that of databases such as ImageNet.Therefore,this paper proposes an unsupervised domain adaptive model.Domain adaptation is a representative method in transfer learning.The source domain and the target domain share the same features and categories,but the feature distribution is different.In the traditional supervised domain,the adaptive source and target domains are labeled,while the unsupervised domain adaptive target domain is not labeled.Unsupervised domain adaptive uses unlabeled target domain to reduce domain offset by training synthetic source domain and existing target domain data,and adjusts the model by aligning the feature distribution between source domain and target domain,so that it can be extended to target data set.The model uses the domain adaptation of the encoder-decoder framework to integrate the adversarial strategy,which is composed of the identifier and the domain classifier.With the help of the gradient flip layer,the adversarial learning is carried out to realize the domain adaptation process.In this paper,experiments are carried out on the off-line handwriting English words and the off-line handwritten Chinese text line data set,respectively.It is proved that the proposed model can achieve the recognition accuracy comparable to the traditional supervised method without any expensive and time-consuming manual marking operation.