| Text recognition technology aims to understand text content from images.With the development of deep learning,text recognition technology has greatly improved to congnize text information.On this basis,text recognition developes some new research directions,including scene text recognition,ancient text recognition,small language text recognition,and damaged text recognition.The damaged text recognition requires a lot of pairs of data,but there is not enough damaged data in actual situation,and it is difficult to simulate the real damaged situation with synthetic data constructed by adding masks to complete text images.In this paper,in response to the insufficient image data of incomplete text,we use Chinese character code to represent the internal information of Chinese character,and directly recognize the damaged character based on the graph neural network.The main research contents of this paper are as follows:(1)a component-code dataset construction using internal information of characters.Damaged character recognition using internal information of Chinese character faces the situation of insufficient data.To address this problem,this paper develops a platform based on the Hierarchical decomposition embedding(HDE)coding method.This platform is used to obtain the code corresponding to the first-level Chinese characters commonly used in modern Chinese.HDE method can make full use of the internal information of Chinese character while ensuring the uniqueness of Chinese character code.The complete Chinese character image of all character classes corresponding to the code are generated with the help of simplified kaiti Chinese character srtyle.On the basis of the complete Chinese character image,the image of(1)damaged character is generated by missing one component.Finally,Chinese character code,integer Chinese character image and damaged Chinese character images constitute the component-code dataset.(2)a damaged text recognition model based on graph neural network.The existing damaged text recognition methods focus on character image restoration,but the restoration methods face to contradiction between a large number of pairs of data demand and the actual damaged data inadequatation.In the damaged character image recognization,the problem of insufficient data still exists.According to the zero-shot and few-shot text recognition methods,the damaged text can be recognized with a small amount of data using known components and a priori knowledge of Chinese character decomposition.In this paper,DynamicGrape V1.0,a damaged text recognition model based on graph neural network,is proposed.DynamicGrape V1.0 uses GraphSAGE with adding edge embedding to optimize the bipartite graph data,which is constructed by Chinese character code to represent the internal information of Chinese character.DynamicGraoe V1.0 use the optimized bipartite graph data to recover the code of damaged Chinese characters.The experiments show that DynamicGrape V1.0 is not useful for both code recovery and character recognition.Therefore,this method needs the further optimization by another way.(3)a damaged text recognition model based on component retrieval.DynamicGrape V1.0 exposes two problems.The first one is that the spaces of the vector which inputs to graph are different,resulting in inconsistent edge data representation.The second one is that the recovery code predicted by DynamicGrape V1.0 is too dense.In order to slove these two problems,this paper proposes DynamicGrape V2.0,a recognition model for damaged text based on component retrieval.The damaged determination module can discriminate complete character and damaged character.The bidirectional embedding transfer module can unify the vector space.The component retrieval strategy can sparse the predictive code and narrow the range of character candidates.Experiments show that DynamicGrape V2.0 has strong recovery ability for code,and this method is able to accurately identify damaged character.Compared with other representative recognition methods,our method achieves better performance. |
