Research Of Number Sequence Recognition Algorithm Based On Transfer Learning

As one of the most concerned research problems in the field of image processing,number recognition technology realizes the digitization of number by detecting,analyzing and recognizing the characters in the image,and is among text recognition technology.Number recognition technology is widely used in daily life and industrial production.The traditional machine learning method to deal with number recognition usually gets the location of the number in the picture first,and then splits out the single character,and then send these segemented characters into the character classifier for recognition.With the improvement of computer performance,deep learning technology has gained more and more attention at present,and the use of deep learning technology to solve the problem of number recognition can overcome many drawbacks in traditional machine learning methods,such as the difficulty to split the adhension of character,or the difficulty to define feature space,etc.It is even possible to realize the ideal process of put in the picture with number,and get the recognition result directly,as an end-to-end method.Currently,there are two main end-to-end technologies for number sequence recognition,one is the attention mechanism and the other is the CTC mechanism.However,deep learning often requires a large number of tagged data samples to build the model.And if the amount of data samples are insufficient,it may cause the results of the overfitting,which leds to the difficulty to ensure the accuracy of the recognition model.So transfer learning will probably help build number recognition model with deep learning.In this paper,a deep learning model based on CTC loss is proposed on the basis of sufficient investigation on traditional methods of number recognition and deep learning.The model can realize end-to-end number sequence recognition and experiments are carried out on large-scale data sets,which verify the good performance of the model,and the feasibility of improving model performance through transfer learning is studied.And then based on the real data set,a large number of experiments prove that transfer learning can improve the adaptive ability of the model,greatly improve the generalization ability of the model,so that it can be widely used in different scenarios.