Research On The Application Of Curriculum Learning In Scene Character Recognition Task

Scene text recognition has always been a challenging problem.Understanding the content of scene text has huge application value in the fields of transportation,social networking,positioning,etc.,and it has become a hot spot on current research.However,scene text recognition is different from traditional document text recognition,scene text representation is rich,and so many problems such as occlusion,bending,and artistic characters seriously affect the accurate of extraction character features by the model,and the recognition accuracy is always unsatisfactory.In order to obtain a deep learning model with a higher character recognition rate,people adopt methods such as changing the network structure and the loss function of the model,correcting distorted characters of the picture,and so on.However,unlike previous work,this article adopts a curriculum learning method that has achieved significant results in classification problems,curriculum learning changes the original method of randomly selecting samples from the dataset,and trains the model in a sample sequence with increased difficulty,which improves the current mainstream scene text recognition algorithm result.Compared with the traditional training method,using the curriculum learning method to train the scenes text recognition model can make the model learn more accurate features at the early stage of training,accelerate the model's convergence speed,and finally better character recognition accuracy is obtained on the test dataset.Using the curriculum learning method,two problems need to be solved: how to define the difficulty of the dataset samples,and how to train the model after obtaining the difficulty of the dataset samples.Around these two key issues,the main works of this paper are as follows:(1)Proposed a method to define the difficulty of scene text and pictures using the key features in the human perspective that cause picture difficulties.First this method preprocesses the Synth Text80 K,ICDAR2013,and ICDAR2015 datasets to unify the picture format.Second uses these processed data to pre-train the model,and then use these pre-trained models to identify the COCO-Text training set data,define the difficulty of the data according to the model recognition scores,and divide the samples of different training subsets based on the these difficulty scores.Use CRNN and ASTER two mainstream scene text recognition algorithms to train on different training subsets according to the order of increasingdifficulty,and finally improve the accuracy of character recognition.Analysis of the experimental results is shown that the method of curriculum learning can effectively improve the convergence rate of the model in the early stage of model training.Train the ASTER and CRNN algorithms in order of increasing difficulty on the COCO-Text dataset to get a maximum increase in 1.8% and 1.13% compared to the benchmark.Training in order of decreasing difficulty reduces the accuracy of the model to vary degrees,and reduces the convergence rate of the model at the initial stage of training.By changing the number of training set subsets and the number of training set subsets samples,we can continue to increase the weight difference between simple and difficult samples,so that the model can find a better local optimal in a certain range.(2)Proposed a method to use key features to define the difficulty of scene text pictures.Using low-level features such as the picture size and the number of label characters to judge the difficulty of pictures in COCO-Text training set.Use the neural network to extract the middle-level feature information such as: the number of objects in the picture,the number of edges,and the number of segments.Link the V-SVR machine learning classifier,fit the information marked by humans to obtain a new difficulty discrimination model,and predicting the difficulty of the COCO-Text training set pictures.Then,similar to the previous step,the curriculum learning algorithm is used to train the model in order of increasing difficulty.According to the experimental results,it can be seen that using the method of this chapter to define the sample difficulty and conducting curriculum learning training can make the CRNN model get a maximum of 0.57% improvement on the COCO-Text dataset.The improvement in the accuracy of character recognition is less than the previous method,but the time loss is greatly reduced.