| The research object of early text recognition technology is scanned text,which is a single,neat text data.With the variety of application requirements,the research object of text recognition has begun to turn to text in daily scenes.Scene text is mainly divided into two categories: The first category is rule text,which distribution is horizontal;The second category is irregular text,which distribution can be of any shape.Although the recognition of regular text can be regarded as a one-dimensional sequence recognition problem,the results obtained by processing irregular text in this way are often not ideal.Therefore,this paper uses two methods to solve the problem.The two methods are the use of two-dimensional attention mechanism and the use of correction method.The first method is to extend the existing one-dimensional attention mechanism and directly decode the two-dimensional feature map through the two-dimensional attention mechanism.This saves the process of compressing two-dimensional information into one-dimensional information,thereby avoiding the introduction of noise and loss of information.This article extends the existing one-dimensional attention mechanism model and introduces a two-dimensional attention mechanism.At the same time,in order to obtain better semantic information expression,the original Long-Short Time Memory-LSTM Network in the model sequence modeling stage was replaced with a Transformer with two-dimensional position coding.Experimental results show that the effect of Transformer is better than LSTM.Through the analysis of the experimental results,a major disadvantage of the model is that the inference time is too long.This is because the operations in the attention mechanism in LSTM are too time-consuming.The accumulated time spent in the recursive operation makes the overall running speed of the model too slow.Therefore,a completely parallel model structure is proposed later and then the two-dimensional attention mechanism is introduced into this model.In this way,the recursive operation structure of LSTM can be discarded and the inference time of the two-dimensional attention model is greatly shortened.The second method is to correct irregular picture text to a horizontal distribution by correcting means,and then make predictions through a one-dimensional sequence recognition network.The core of the method is how to correct the text on the picture.This can be done through a spatial transformation network.The spatial transformation network uses a two-dimensional interpolation algorithm to generate corrected text images.After that,the corrected text image is input into the convolutional neural network.This process is called feature extraction.Then the extracted features is input into the Bilateral Long-Short Time Memory Network for sequence modeling.Finally,a prediction network is used to output the results.This paper compares the effects of two different prediction networks through experiments.The two prediction networks are CTC network and one-dimensional attention mechanism LSTM network.The experimental results show that these two prediction networks have their own advantages and disadvantages.Compared with the two-dimensional attention method,the correction method has an advantage in accuracy but a disadvantage in speed.In order to find out which factors are likely to cause identification errors,we classify the examples of identification errors and summarize several common causes of identification errors. |
